VIEW THIS AS

Auto mode follows the Route Engine until you choose a viewpoint.

YOU ARE HERE

ROUTE CHECK

CONNECTED TO

WHAT NEXT

Use the canonical route for this room, or HELP if you are unsure.

Super Intelligence | The Environment: Climate, Biodiversity and Competing Objectives

Three secondary students review open books and notes together in a classroom.

Super Intelligence master guide › What is Super Intelligence? › The environment

Three students studying together with open books at a classroom table.
An illustration of careful study from the eduKate Singapore image collection. It is not evidence from an environmental field project.

Environment · Evidence literacy

Learn to check environmental AI claims, reproduce the calculations, and keep climate, biodiversity, resources and legitimate decisions distinct.

Artificial intelligence can help people interpret environmental measurements, compare forecasts and explore possible responses. It cannot turn an incomplete measurement into a complete picture of nature, turn a prediction into proof of a cause, or decide whose losses are acceptable merely by producing an impressive score. A useful environmental AI claim has to survive several different questions: what was measured, what was predicted, what changed because of the intervention, what resources the system used, and who had the authority to choose the outcome.

This guide is for students, parents, teachers and interested readers who want to examine those questions without needing to build a climate model. Its purpose is environmental evidence literacy. You will work through complete fictional examples of habitat screening, temperature prediction, electricity comparisons, computing footprints and competing objectives. The arithmetic is deliberately small enough to check independently. None of the fictional numbers is a measurement from Singapore, an estimate of a real product, a carbon credit, or an instruction to alter a habitat or public service.

In this series, practical “Super Intelligence”, or SI, is an editorial umbrella for current AI tools and the human systems around them. Artificial superintelligence, or ASI, means a hypothetical system with broadly superhuman capabilities. The environmental applications discussed here do not establish that ASI exists. A better forecast in a bounded test is evidence about that forecast, not proof of general wisdom, universal expertise or a right to govern.

The central habit is to keep claims separate until the evidence justifies connecting them. Better classification might improve the information available to an ecologist. Better information might support a different decision. That decision might improve an environmental outcome. Each “might” names a further question, not a step that can be silently skipped. Conversely, a computing footprint does not by itself tell us whether an entire application has a positive or negative environmental effect. Benefits, burdens, alternatives, time and place all need an explicit comparison.

Evidence and source availability were checked on 1 October 2026. Published assessments are dated below; older sources are retained where they establish the relevant scientific distinction. New measurements can update a conclusion without changing the discipline used to reach it.

1. Separate five questions before evaluating an environmental claim

Observation asks what was recorded. A sensor reports a temperature, a camera records an animal, or a satellite instrument measures radiation that a processing system converts into a land-cover estimate. The observation includes an instrument, location, time, method and quality checks. “The map contains 800 green pixels” is an observation about the processed map. “The site contains healthy native habitat” is a further interpretation. It requires knowledge about what those pixels represent and what the map cannot see.

Prediction asks what a model expects under stated conditions. It may estimate tomorrow's temperature, the chance that a surveyed location contains a species, or the electricity a building would use under a specified schedule. Predictions can be useful even when the model does not identify the underlying causal mechanism. A strong test checks performance on genuinely new cases, compares it with a relevant alternative and examines consequential mistakes rather than only an average score.

Attribution asks what caused a change or how much a factor contributed to it. If electricity use falls after an AI system is introduced, the timing alone does not establish that the AI caused the fall. Weather, occupancy, repairs, prices and measurement changes may also matter. Attribution needs a defensible comparison with what would have happened without the intervention. That comparison is called a counterfactual. It is usually estimated, so its assumptions belong in the conclusion.

Optimisation asks which choice scores best under an objective and constraints. A programme can search schedules to minimise electricity use while keeping a specified temperature range. It can compare conservation options under a budget. The mathematical result is conditional on the supplied objectives, feasible choices, predicted effects and constraints. If the objective ignores habitat quality or access for residents, the optimiser does not magically include those concerns because its search is sophisticated.

Legitimate decision-making asks who may choose, through what process, with what obligations to people and places. It involves rights, expertise, participation, reasons, accountability and sometimes formal legal procedures. An accurate map does not grant permission to change land use. A technically efficient plan does not establish that affected communities accepted its burdens. AI can assist the evidence and explanation, but its output cannot replace the decision process.

Consider the sentence, “Our AI restored the wetland.” It compresses all five questions. Did the system classify images more accurately? Did a forecast guide work? Was restoration actually undertaken? Which ecological indicators changed, over what period, relative to which comparison? Were adverse effects checked elsewhere? Who approved the work? Until those answers exist, a narrower statement may be warranted: “The system flagged possible vegetation changes for expert review.” The narrower sentence is more useful because readers can identify what was accomplished.

The five questions also prevent a familiar reversal: assuming that uncertainty about one local intervention undermines established climate science. A weak evaluation of a building-control pilot says little about the global evidence for human-caused warming. The scale, evidence and claim differ. Likewise, a strong global assessment cannot certify an untested local product. Evaluate each claim at the level where its evidence actually operates.

A practical reading method is to underline the verb in a claim. “Detected”, “forecast”, “caused”, “maximised” and “approved” require different evidence. If a paragraph moves from one verb to another, look for the missing bridge. A report may legitimately contain all five stages, but it should not present them as interchangeable synonyms for progress.

Back to contents · Continue: 2. Read current environmental evidence without borrowing its authority

2. Read current environmental evidence without borrowing its authority

The IPCC's 2023 Synthesis Report brings together its sixth assessment cycle. It distinguishes mitigation, which addresses emissions and removals, from adaptation, which addresses impacts and risks. It also discusses maladaptation and the importance of inclusive, locally appropriate approaches. These findings support careful consideration of both physical outcomes and the processes used to select responses. They do not certify any AI vendor or particular local intervention. IPCC, AR6 Synthesis Report, 2023.

For a more recent observation, WMO's State of the Global Climate 2025, published on 23 March 2026, reports that 2015–2025 were the hottest eleven years on record and places 2025 about 1.43°C above the 1850–1900 average. That statement concerns a particular annual global temperature estimate and reference period. It is not a local afternoon temperature, a forecast for a specific neighbourhood, or a measurement of an AI project's contribution. WMO, State of the Global Climate 2025.

The 2024 IPBES Nexus Assessment examines interconnections among biodiversity, water, food, health and climate change. It provides a reason to examine connected effects rather than treating a gain on one environmental dimension as a complete verdict. Separately, IPBES's 2026 Business and Biodiversity Assessment addresses the measurement of business impacts and dependencies. A measurement method needs to fit the decision and context; the existence of a sophisticated indicator does not make every application of it appropriate. IPBES Nexus Assessment and IPBES Business and Biodiversity Assessment.

These are assessment sources, not inputs to the fictional calculations in this guide. Their findings do not establish the invented error rates, electricity factors or habitat scores used below. Keeping that boundary visible matters. A real citation placed beside invented numbers can make a practice exercise look like a documented field study unless the relationship is explained.

When an AI assistant summarises an assessment, ask it to identify the report edition, chapter or summary section, claim and limitation. Then inspect the source. Check whether the source speaks about observed change, projected change or a conditional policy pathway. A scenario is not automatically a forecast of what will happen. A statement about a global average is not automatically true at every location. A range is not an invitation to quote only the most convenient endpoint.

Also distinguish the observation year from the publication year. A report released in 2026 may primarily describe conditions in 2025. A web page can have a later update date without changing the reference period of its headline statistic. Recording all three dates, when relevant, prevents an apparently current dashboard from silently mixing observations of different ages.

A trustworthy environmental explanation does not need a dramatic global number in every paragraph. Often the most useful information is narrower: which plots were sampled, how missing data were handled, whether the comparator received equivalent maintenance, and what remains unknown. Prestigious sources provide context. Local evidence must still carry the local claim.

Back to contents · Continue: 3. Use units and denominators that match the question

3. Use units and denominators that match the question

Energy, power, emissions, water use and ecological condition are different quantities. Power describes a rate of energy use. Energy records use over time. If fictional equipment draws a constant 2 kilowatts for 5 hours, it uses 10 kilowatt-hours. The multiplication is 2 kW × 5 h = 10 kWh. A report that compares 2 kW with 10 kWh as though one were simply larger has lost the time dimension.

To estimate electricity-related emissions, a calculation may multiply electricity consumption by an emissions factor with matching units. In a deliberately simplified example, 10 kWh × 0.4 kg CO2e/kWh = 4 kg CO2e. The kilowatt-hour unit cancels. The factor here is invented for teaching. A real calculation must identify the factor's source, year, geography and accounting purpose. A factor appropriate for an annual inventory may not describe the emissions change caused by moving one activity to a different hour.

CO2e means carbon-dioxide equivalent under a stated comparison method. It is not a direct measurement of every ecological effect. Nor is a tonne of CO2e a hectare of habitat, a litre of water, or a measure of community wellbeing. Converting money into a common currency does not imply that every environmental value can be converted into a scientifically settled single currency. Some comparisons need several indicators displayed together.

Stocks and flows also need separation. A stock describes an amount at a time, such as an estimated quantity of carbon stored. A flow describes a change or transfer over a period, such as annual emissions. An estimated stock of 100 tonnes cannot simply be subtracted from 10 tonnes of annual emissions each year indefinitely. The accounting would repeatedly spend the same stock. Questions about permanence, additional change, measurement and loss remain even when the arithmetic looks neat.

A denominator tells the reader “per what”. Electricity per processed image may fall while total electricity rises because the service processes more images. Detections per survey visit may rise because surveys moved to easier sites. Emissions per product may fall while total production grows. Useful reporting gives both the rate and its total: accepted reports, attempted reports, kilowatt-hours, hours of service, area surveyed and missing observations as appropriate.

Biodiversity requires particular care. A count of species, their abundance, the condition of habitats and their connectivity describe different features. Ten common species recorded beside a path do not make that location interchangeable with another containing an ecologically important population. A tree count does not establish species composition, survival, habitat suitability or long-term condition. Counting what is easy can be useful, provided it is not renamed as a broader outcome.

Spatial resolution is another denominator in disguise. A one-kilometre grid cell averages over a large area. A fine-looking display may interpolate those values into tiny coloured squares without creating new measurements. Ask for the original measurement scale, not merely the display resolution. A model can produce a prediction for a small place, but that prediction needs validation at a suitable scale before being treated as a local observation.

A good comparison states its functional unit: the service being compared. For the worked computing example later, that unit is one accepted environmental reading packet meeting a fixed rubric. We will still show all attempted packets and total energy. This avoids rewarding a system that produces many unusable drafts cheaply or penalising one simply for making fewer attempts. Defining the service before examining results is a safeguard against changing the denominator to favour a preferred answer.

Back to contents · Continue: 4. Packet A: evaluate an observation system without inventing an ecosystem result

4. Packet A: evaluate an observation system without inventing an ecosystem result

Everything in Packet A is fictional. A class evaluates two systems that screen existing, authorised photographs for a visible plant category. The work is desk-based. No student enters a habitat, collects specimens or publishes species locations. Independent reviewers have agreed on reference labels for 200 photographs: 40 contain the target category and 160 do not. The set comes from several locations and dates, but its small size is not claimed to represent any region.

System Reed flags 50 photographs. Of those, 32 contain the target and 18 do not. Among the 150 unflagged photographs, 8 contain the target and 142 do not. System Fern flags 60 photographs. Of those, 36 contain the target and 24 do not. Among its 140 unflagged photographs, 4 contain the target and 136 do not. These four counts per system are the full confusion matrices; no unreported cases are omitted.

Reed's precision is 32 true detections divided by 50 flags, or 64%. Its recall is 32 divided by all 40 target photographs, or 80%. Fern's precision is 36/60 = 60%; its recall is 36/40 = 90%. Fern finds four more of the target photographs, at the price of ten additional flags requiring review, six of which are false positives. Whether that is useful depends on the purpose and review capacity.

Overall accuracy gives another view. Reed gets 32 + 142 = 174 photographs correct, so accuracy is 174/200 = 87%. Fern gets 36 + 136 = 172 correct, or 86%. A headline using only accuracy would favour Reed. A screening task where missed targets are especially costly might favour Fern, provided reviewers can check all flags. Neither conclusion follows from a universal rule that one metric is always the best.

There is an instructive baseline. A system that labels every photograph “target absent” gets all 160 negatives correct and none of the 40 positives correct. Its accuracy is 160/200 = 80%, but recall is zero. That baseline exposes why a seemingly high accuracy can coexist with complete failure at the intended screening job. Precision for a system issuing no positive flags is undefined here, because it has no positive predictions; it should not be printed as a meaningful perfect score.

Now suppose the class has capacity to review only 50 flags. Reed fits that capacity; Fern does not. We cannot infer Fern's performance on a selected top 50 because the packet gives no ranked confidence scores. Removing ten arbitrary flags could remove true detections. The correct next step is to request a pre-specified ranking test or increase review capacity, not assume that Fern retains its 90% recall after trimming its output.

The comparison also needs a quality check on the reference labels. Reviewers may disagree about a blurry photograph. For this exercise the reference labels are stipulated, but a real evaluation should report the labelling protocol, adjudication process and unresolved cases. A reference label is a carefully constructed comparison point, not a guarantee that humans never make mistakes.

Most importantly, the target is a visible category in photographs. We have not estimated population size, habitat health, reproductive success or restoration impact. Several photographs could show the same organism; absent photographs do not establish absence from the site. Detections could change because camera placement, lighting or season changed. The justified conclusion is therefore about screening performance on this packet, not “Fern increased biodiversity”.

A useful repair to the evaluation would reserve photographs from entirely separate sites and later dates for testing. If near-identical images appear in both development and test sets, apparent performance can reflect familiarity with the scene. Also report errors by lighting, background and image quality, where sample sizes permit. A single average may conceal a category of cases where the system systematically misses the target.

The completed decision note is short: “Fern found 36 of 40 reference-positive photographs, compared with Reed's 32, but generated 60 rather than 50 review flags. Our current capacity is 50. We will not claim an ecological outcome or deploy automatic field action. We need an independent ranked-screening evaluation or a larger review budget before choosing a workflow.” The note states what is known and what decision remains open.

Back to contents · Continue: 5. Packet B: distinguish forecast skill from climate explanation

5. Packet B: distinguish forecast skill from climate explanation

Packet B is also fictional. An education group compares two models for a daily maximum-temperature prediction at one imaginary monitoring point. All temperatures are in degrees Celsius, all predictions are issued before the relevant day, and both models receive the same information cutoff. The four observed values, in date order, are 30, 32, 34 and 32. Model Moss predicts 31, 31, 33 and 33. Model Cloud predicts 30, 32, 37 and 31. The simple reference forecast predicts 32 on every day.

Mean absolute error, or MAE, averages the size of the error without its sign. Moss's absolute errors are 1, 1, 1 and 1°C, giving an MAE of 4/4 = 1°C. Cloud's are 0, 0, 3 and 1°C, also giving 4/4 = 1°C. The reference errors are 2, 0, 2 and 0°C, again giving 4/4 = 1°C. On this tiny test, neither model improves this metric over the simple reference.

Mean error tells a different story because positive and negative errors can cancel. Using prediction minus observation, Moss has errors +1, −1, −1 and +1°C, averaging zero. Cloud has 0, 0, +3 and −1°C, averaging +0.5°C. A zero mean error does not mean Moss was exact. It means the signed errors balanced on these four cases. A report using only mean error would hide the fact that every Moss prediction was wrong by one degree.

Suppose a purely classroom label marks an observed maximum of 34°C or higher as “threshold reached”. This is an invented exercise threshold, not a health warning or an operational recommendation. Only day three reaches it. Moss predicts no threshold days, so it misses the one positive. Cloud predicts day three above the threshold and matches that classification, although its predicted temperature is three degrees too high. The reference also misses day three.

This reveals why the relevant evaluation must follow the intended use. If the classroom task is numerical prediction, MAE and bias matter. If the task is a threshold classification exercise, missed and false threshold events matter too. We cannot select the winning metric after seeing the results without reporting that choice. And with only one threshold event, a success or failure is far too little evidence for a stable estimate of future performance.

Now imagine a separate fictional probability check. Across twenty independent classroom cases where a model assigned a 70% chance of the defined event, the event occurred fourteen times. Fourteen divided by twenty is 70%, so this small group is consistent with the stated probability. It does not prove perfect calibration. We would need more cases, other probability bands, a fixed event definition and a check that observations are sufficiently independent for the analysis being used.

A probability interval or ensemble range also needs interpretation. A display labelled “80% interval” should explain how that interval was constructed and what coverage it achieved on held-out cases. The fact that a model outputs several plausible futures does not itself prove that its spread represents all relevant uncertainty. Missing mechanisms, measurement bias and unusual conditions can affect every member of the ensemble together.

Climate projections introduce another distinction. A long-term projection under a specified emissions pathway asks what conditions may follow if that pathway and modelling assumptions apply. A short-term weather forecast tries to predict a near-term state given current information. They overlap scientifically but are not interchangeable evaluation tasks. Four days at one monitoring point cannot validate a multi-decade climate projection, and a successful climate-scale explanation cannot promise tomorrow's exact local temperature.

Nor does forecast skill establish attribution. Suppose hotter days coincide with lower park attendance. A model might predict attendance well from temperature. That alone does not establish how much an intervention changing shade would alter attendance, because schedules, rainfall, holidays and other factors could affect both the input pattern and the outcome. Predicting under the current arrangement differs from predicting what happens after changing that arrangement.

The completed Packet B report should therefore say: “Both AI models and the constant reference have an MAE of 1°C on four fictional cases. Cloud identifies the only exercise-threshold day but overpredicts its temperature. The sample is inadequate for deployment claims. We need a longer, pre-specified test including relevant seasons and error categories.” It should not say that the model “understands climate change” or is ready to issue public warnings.

For independent practice, cover the calculations and rebuild them from the four input rows. Check that each denominator is four for a daily mean, one for recall on the single positive threshold day, and twenty for the separate probability group. The ability to recover those denominators is more valuable than memorising the name of the best-looking model.

Back to contents · Continue: 6. Packet C: ask whether an intervention caused a reduction

6. Packet C: ask whether an intervention caused a reduction

Packet C concerns a fictional desk review of two buildings. Students do not change any real controls. Building A uses 12,000 kWh during a four-week baseline and 10,200 kWh during a four-week follow-up after an AI-assisted scheduling pilot. Building B, a comparison building without the pilot, uses 10,000 kWh at baseline and 9,500 kWh at follow-up. Both periods contain the same number of days. The packet initially supplies no weather, occupancy or maintenance information.

The observed reduction in A is 12,000 − 10,200 = 1,800 kWh, or 15% of its baseline. That is a before-and-after result. It is not yet a causal estimate. B also falls by 500 kWh, or 5%. The fact that a comparison building changed suggests that conditions other than the pilot may matter, but it does not tell us exactly how much they matter in A.

One illustrative counterfactual assumes that without the pilot A would have experienced the same proportional change as B. B's follow-up-to-baseline ratio is 9,500/10,000 = 0.95. Applying it to A gives an estimated counterfactual of 12,000 × 0.95 = 11,400 kWh. Compared with A's observed 10,200 kWh, the modelled difference is 1,200 kWh. It is 10% of A's baseline, and about 10.53% of the 11,400 kWh counterfactual. Both percentages are correct if their denominators are named.

Another illustrative method assumes the same absolute change rather than the same proportional change. Under that assumption, A's counterfactual is 12,000 − 500 = 11,500 kWh, making the modelled difference 1,300 kWh. The two estimates disagree because the assumptions differ. The packet provides no reason to declare one assumption true. Reporting only the larger estimate would conceal a choice in the method.

Additional information now arrives: A had fewer occupied hours during follow-up, while B's schedule was unchanged. That weakens both simple comparisons, because A and B may not share the same untreated trend. We cannot fix this by declaring the difference “small” without data. We need the occupied-hour records, relevant weather measures, operating conditions, changes in equipment and a method suited to those differences.

The word “adjusted” is not a guarantee either. An analyst can adjust for recorded variables while important unrecorded differences remain. Some variables may themselves be affected by the intervention; controlling for them can change the question being estimated. A careful evaluation specifies the causal question first, then justifies its design and adjustment choices. A classroom reviewer can identify missing evidence without pretending to settle the appropriate statistical model.

For teaching arithmetic only, suppose a clearly stated scenario uses the 1,200 kWh modelled difference and an invented constant factor of 0.4 kg CO2e/kWh. Multiplication gives 480 kg CO2e, or 0.48 tonnes, because 1,000 kilograms make one tonne. This is a scenario-derived electricity-related estimate under the supplied assumptions. It is not a verified emissions reduction, a lifecycle assessment or an offset. Hardware, computing, maintenance and other changes have not been included.

The service must also be comparable. A lower electricity meter reading achieved by leaving rooms unusable is not an equivalent service improvement. A real evaluation would define comfort, safety, accessibility, availability and maintenance requirements with qualified operators. It would record violations rather than averaging them away. Energy is an objective within a service, not a licence to remove the service.

A stronger study could use a pre-specified design with suitable comparison periods or groups, enough observations and independent review. Randomised assignment may help where feasible and appropriate, but it is not permission to experiment on people's essential services without oversight. An evaluation design should match both the causal question and the risks of altering the system. For students, existing anonymised data and fictional packets are enough to learn the reasoning.

The final claim is therefore bounded: “A used 1,800 kWh less during follow-up. A simple proportional-comparator scenario attributes a 1,200 kWh difference to the pilot, but occupancy changes undermine that attribution. Further evaluation is needed before a causal savings claim.” This preserves the observed fact while resisting an unjustified explanation. It also makes the next evidence request specific enough to act on.

Back to contents · Continue: 7. Packet D: compare climate and biodiversity objectives without hiding a value choice

7. Packet D: compare climate and biodiversity objectives without hiding a value choice

Packet D is a fictional planning exercise involving three designs for an imaginary educational garden. It authorises no planting, construction or spending. All figures are made-up scenario outputs over the same stated year. The objective is to practise comparison, not to estimate how real trees, roofs or habitats perform. Each design has already been assumed feasible for the arithmetic, but real feasibility and permissions would require separate work.

Option Canopy has an estimated electricity-related emissions reduction of 1.8 tonnes CO2e, a habitat-condition teaching index of 40 points, a cost of 8 budget units and 200 hours of shaded access. Option Mosaic has 1.2 tonnes, 80 points, 9 budget units and 180 shaded hours. Option Courtyard has 1.0 tonne, 30 points, 10 budget units and 150 shaded hours. Higher is better for the three benefits; lower is better for cost. The habitat index is an invented classroom score, not an offset currency or validated ecological metric.

On these stipulated values, Courtyard is dominated by Canopy: Canopy has more of every listed benefit and a lower cost. Mosaic also improves every listed benefit over Courtyard while costing less. A dominated option can usually be set aside within the stated comparison, unless an omitted consideration changes the situation. The qualification matters. We have not assessed every possible feature simply because we assessed four.

Canopy and Mosaic are not ordered so easily. Canopy offers more estimated emissions reduction, lower cost and more shaded hours. Mosaic offers a higher habitat index. Neither dominates the other. These are competing options on the stated dimensions. An optimisation system that returns only one without explaining its objective has hidden the very disagreement readers need to see.

Suppose the exercise first uses a weighted score for just the emissions and habitat dimensions. To avoid mixing raw units, divide the emissions values by the fixed reference of 2 tonnes and habitat points by 100. Canopy then has normalised values 0.9 and 0.4; Mosaic has 0.6 and 0.8. These references are chosen for the exercise, not supplied by nature. The choice of scale is part of the decision model.

With 70% weight on emissions and 30% on habitat, Canopy scores 0.7 × 0.9 + 0.3 × 0.4 = 0.75. Mosaic scores 0.7 × 0.6 + 0.3 × 0.8 = 0.66. Canopy wins this score. Reverse the emphasis to 30% emissions and 70% habitat: Canopy scores 0.55 and Mosaic 0.74. Mosaic now wins. The model has not contradicted itself; the priorities changed.

Let the emissions weight be w and the habitat weight be 1 − w. Canopy's score becomes 0.4 + 0.5w, while Mosaic's becomes 0.8 − 0.2w. They tie when 0.7w = 0.4, so w = 4/7, about 57.14%. Above that emissions weight Canopy leads; below it Mosaic leads. Showing the switching point is more informative than announcing a single “optimal green design”.

Now change the structure from weights to constraints. Suppose the legitimate decision-makers require at least 150 shaded hours, a cost of no more than 9 units and a habitat index of at least 60 for this exercise. Mosaic meets all three. Canopy fails the habitat threshold; Courtyard fails both the cost and habitat thresholds. The result comes from explicit requirements rather than a compensation rule allowing one benefit to cancel another deficit.

This is especially important where some matters should not be traded away. Rights, protected features and safety obligations may operate as hard constraints, subject to the relevant legitimate process. A weighted score that lets cheapness compensate for an unacceptable violation has encoded the wrong decision, even if calculated perfectly. The exercise's numerical thresholds are not statements of actual law or universal ecological standards.

Uncertainty may change the apparent ranking. Imagine the estimated emissions reductions have ranges of 1.2–2.2 tonnes for Canopy and 0.8–1.6 for Mosaic. These invented ranges overlap. The midpoint ranking alone does not establish which design performs better in reality, and interval overlap alone does not settle the statistical comparison either. We need to know what the ranges mean, whether errors are related and how estimates were constructed.

A useful decision note therefore gives the options, dominance check, constraints, weights and sensitivity. It might recommend collecting better evidence about habitat condition before committing to an irreversible design. It might identify a reversible trial of a non-invasive measurement method. It should not call a classroom index “biodiversity gained” or claim that a tonne of avoided emissions compensates for damage to an irreplaceable ecological feature.

The deeper lesson is that an optimiser answers a question someone has framed. More computational power can expand the search, expose compromises and test robustness. It cannot establish the legitimate weights on its own. Readers should ask whether the model made the value choice visible and whether affected people had a meaningful route to challenge it.

Back to contents · Continue: 8. Packet E: compare computing footprints per useful result

8. Packet E: compare computing footprints per useful result

An environmental AI system uses physical resources as well as information. UNEP's 2024 issue note calls for assessing environmental effects across the AI lifecycle, including hardware and software. This supports examining manufacturing, operation and end-of-life rather than treating one measured inference run as the whole footprint. The note does not supply the invented factors or workflow figures used in the following packet. UNEP, Artificial Intelligence end-to-end, 2024.

Packet E compares two fictional ways to prepare exactly the same kind of environmental reading report. Workflow Oak attempts 1,000 reports and produces 800 that pass a fixed independent rubric. Its total measured IT electricity for those attempts is stipulated to be 40 kWh. Workflow Willow attempts 1,000 reports and produces 900 accepted reports, using 30 kWh of IT electricity. The rubric covers factual traceability, complete units and explicit limitations; a fluent but unsupported report fails.

For this exercise, a measured facility-to-IT energy ratio of 1.5 is assumed for both workflows over the same period. Multiplying gives 60 kWh of allocated facility electricity for Oak and 45 kWh for Willow. The ratio is used here as an allocation assumption. If a real measurement already includes facility overhead, multiplying it again would double-count that overhead. The boundary has to be established before using the ratio.

Using the same invented factor of 0.4 kg CO2e/kWh gives 24 kg CO2e of electricity-related emissions for Oak and 18 kg for Willow. Divide by accepted reports: Oak is 24/800 = 0.03 kg, or 30 grams, per accepted report. Willow is 18/900 = 0.02 kg, or 20 grams. On this stipulated operational boundary, Willow uses less electricity in total and less per accepted report while producing more accepted reports.

A denominator based only on attempts would obscure the quality difference. Oak uses 24 grams per attempt and Willow 18 grams per attempt. Those are valid rates, but they answer a different question. Reporting both attempt and acceptance totals shows how the service is functioning. It also prevents rejected drafts from disappearing from the electricity numerator while remaining in an impressive productivity claim.

Now add a fictional allocated manufacturing burden of 8 kg CO2e for Oak and 18 kg for Willow over this batch. The expanded totals become 32 kg and 36 kg respectively. Oak is 32/800 = 40 grams per accepted report; Willow is 36/900 = 40 grams. The per-accepted-report result now ties, even though Willow's total burden is larger because it supplies more accepted reports. A comparison at equal service volume would need a justified scaling model, not an assumption that every cost scales perfectly.

The manufacturing allocations are stipulated numbers, not direct measurements of marginal harm from a particular request. Real allocation choices involve equipment lifetimes, shared use and what hardware would otherwise have been purchased or kept in service. An inventory allocation helps account for a share of existing burdens. It does not automatically answer the consequential question of what changes if the user sends one more request or replaces a system.

Water requires a separate line. Suppose fictional onsite consumptive water factors are 0.2 litres per allocated facility kWh for Oak and 0.4 litres for Willow. The resulting onsite consumption is 12 litres and 18 litres per batch. That is 0.015 litres per accepted Oak report and 0.02 litres per accepted Willow report. Willow's lower operational electricity-related emissions do not give it a lower value on this particular water measure.

Do not extend that result beyond the packet. Water withdrawal and water consumption are different accounting quantities; water may be returned, consumed or changed in quality. Location, season, source and competing demand matter. Indirect water associated with electricity and equipment manufacture is excluded from this miniature calculation. Adding it would require compatible data and care not to count the same water twice.

The honest conclusion is therefore multidimensional. Willow leads on the supplied operational electricity measures, ties Oak on the expanded allocated emissions per accepted report, and has higher stipulated onsite consumptive water use per accepted report. None of this establishes the total environmental benefit of either application. That requires evidence about what useful action the reports enable, compared with realistic alternatives, and whether that action actually occurs.

A small, reliable model or an ordinary spreadsheet may be enough for a bounded task. A larger AI workflow can be appropriate where it produces demonstrably better accepted work. The comparison should include a non-AI baseline where relevant and should count retries, retrieval, verification and human review consistently. “Use AI” and “avoid AI” are both too coarse to substitute for a service-specific comparison.

Back to contents · Continue: 9. Follow effects beyond the project boundary

9. Follow effects beyond the project boundary

Leakage occurs when an apparent improvement within one boundary is accompanied by displaced activity or harm elsewhere. In an environmental evaluation, it is a question about a wider system. If a fictional programme reduces a reported pressure in its target area while the pressure increases in a neighbouring area because activity moved, the local dashboard may overstate the system-wide improvement. Establishing that displacement is causal still requires evidence; a simultaneous increase elsewhere is a warning to investigate, not proof by itself.

For a transparent arithmetic exercise, suppose a scenario predicts 100 units of avoided pressure inside its boundary and stipulates that 30 units are displaced outside because of the intervention. Under those assumptions, the wider avoided amount is 100 − 30 = 70 units. The displacement rate is 30/100 = 30% of the gross in-boundary amount. The “units” have to describe a compatible quantity before subtraction makes sense. Thirty hectares and one hundred tonnes cannot be combined this way.

Rebound concerns changes in use following an efficiency improvement. Consider a fictional service that initially runs 1,000 tasks at 0.04 kWh each, for 40 kWh. A more efficient version uses 0.03 kWh per task. At the original volume, it would use 30 kWh, saving 10. If use expands to 1,200 tasks, actual use is 36 kWh. The total reduction is now 4 kWh. Six of the ten expected kWh of savings have been taken back by increased volume, a 60% arithmetic rebound relative to that fixed-volume counterfactual.

At 1,600 tasks, the new total is 48 kWh, exceeding the original 40 despite lower energy per task. This is a hypothetical sensitivity calculation, not a prediction that every efficiency improvement increases total demand. To explain an observed increase causally, we would need evidence about why volume changed and what would have happened without the efficiency improvement. Demand might rise for unrelated reasons.

Leakage and rebound require a comparison boundary that is wide enough to catch relevant effects but still practical to evaluate. “Count everything everywhere forever” is not an operational method. A more useful approach names likely pathways, identifies material risks, gives available evidence and states exclusions. A report might examine whether saved staff time led to more computing, whether a supply-chain change moved burdens, or whether reduced costs enabled an environmentally beneficial service to reach more people.

Not every increase in activity is undesirable. More accepted reports, safer services or better monitoring can have value. The point is to stop calling a per-unit efficiency gain an automatic absolute reduction. Present the service change and the resource change separately, then explain the judgement. This lets readers decide whether the additional service is worth its additional burdens under the relevant goals.

Avoided emissions claims need a credible alternative as well. Comparing a new system with an obsolete, unusually wasteful baseline can create a large paper benefit while saying little about the choice actually available. A fair comparator might be a maintained existing system, a simpler improvement or doing less of an unnecessary task. Include the cost of getting an intervention adopted, operated and maintained; an unimplemented recommendation does not produce the scenario's predicted effect.

Offset language needs stronger discipline than an ordinary comparison. An emissions estimate from a classroom calculation is not an offset credit. A habitat score is not evidence that ecological harm has been compensated. Questions of additionality, durability, displacement, verification, equivalence and applicable rules cannot be settled by subtracting two convenient numbers. This guide teaches readers to ask those questions; it does not provide a method for issuing or validating credits.

Back to contents · Continue: 10. Keep local knowledge, rights and affected people inside the decision

10. Keep local knowledge, rights and affected people inside the decision

Environmental evidence is always situated somewhere. A global model can provide useful context while missing a blocked drain, a seasonal use of land, a culturally significant place or a local species interaction. Local knowledge is not merely an anecdote to be replaced once a large dataset arrives. It can identify missing variables, challenge categories and reveal impacts that the chosen measurement system never collected.

The IPCC's 2023 Synthesis Report specifically discusses meaningful participation and the contribution of scientific, Indigenous and local knowledge to locally appropriate climate-resilient development. This is a statement about the value of inclusive processes, not a claim that every disagreement has an easy resolution. Different knowledge sources can complement one another while needing careful interpretation, consent and attention to context. IPCC, Summary for Policymakers, section C.6.

Consider an imaginary map that labels a piece of land “underused”. The label might mean low visible building density in a dataset. People may use the place for access, gathering, informal work or practices invisible to the instrument. An optimiser treating the label as proof of spare land would convert a data limitation into a decision with consequences. The repair is to investigate what the category measures and hear from those affected before defining the feasible options.

Participation should happen early enough to influence the problem, not only after the model has selected a preferred plan. Ask who helped choose the objectives, who can challenge the data, who understands the uncertainties and who can require reconsideration. A public explanation should distinguish the authority that made a decision from the model that supplied an estimate. Naming a software system as the decision-maker can blur responsibility without providing accountability.

Information sharing also has limits. Precise locations of vulnerable species or sensitive community knowledge may need protection. A student should not upload such information to an AI service simply because it would improve a prompt. Obtain appropriate permission and use authorised, suitably aggregated or fictional data. The educational value lies in testing reasoning, not in exposing ecological or personal information.

In a Singapore classroom, the same discipline applies without pretending that a fictional garden packet describes a real neighbourhood. Tropical heat, rainfall, shade, water management and biodiversity can be relevant subjects, but actual conditions and permissions must be checked through appropriate local sources and qualified people. This guide's temperature thresholds, emissions factors and habitat scores are deliberately not operational local advice.

A rights-aware comparison can still be practical. Add a line to every option recording who bears a burden, whether that group was heard, which concerns remain unresolved and who has authority to decide. Some concerns will be represented by measurements; others will be constraints or reasons. Do not force every concern into a numerical score merely to make the spreadsheet uniform.

There will sometimes be genuine conflict: one option may reduce one pressure while worsening another, or distribute benefits unevenly. A transparent process need not promise unanimous agreement. It should make the disagreement legible, justify the choice, preserve routes for challenge and monitor what actually happens. More capable AI could help explore alternatives and explain consequences, but increasing intelligence would not erase the need for legitimacy.

Back to contents · Continue: 11. Build a complete environmental evaluation packet

11. Build a complete environmental evaluation packet

A reusable packet begins with a precise claim and intended use. Instead of “AI for a greener future”, write “screen authorised historical photographs for expert review” or “compare two report-generation workflows under a fixed acceptance rubric”. State who will use the result and what action it may inform. If the action is high consequence, the evaluation and governance requirements become stronger. The classroom examples in this article stay at desk-based analysis.

Next provide a boundary sheet. Record the place, time period, functional unit, included activities, excluded activities and comparator. Mark every number as observed, supplied, estimated, modelled or hypothetical. A measured electricity total and an assumed emissions factor can legitimately appear in the same calculation, but their different evidential status must remain visible. Do not relabel the final product as wholly measured.

Then supply the full inputs needed to reproduce the result. For a classifier, include all four confusion-matrix counts and the sampling protocol. For a forecast, include dates, information cutoffs, predictions, outcomes and the reference forecast. For a footprint, include the energy boundary, quality denominator, factor and allocations. A reader should not need hidden rows to understand why the conclusion follows.

A method sheet explains the arithmetic and analytical choices. Name the denominator before giving a percentage. Describe how missing cases are handled. Identify whether the result is a before-and-after observation, a counterfactual estimate or a scenario comparison. If there are competing methods, show how the conclusion changes under plausible alternatives rather than choosing the preferred method silently.

The outcome sheet distinguishes task performance from environmental outcomes. In Packet A, the outcome is correct screening on supplied photographs; biodiversity improvement has not been measured. In Packet B, the outcome is prediction error on a small test; no climate intervention was evaluated. In Packet C, lower electricity use was observed but causal attribution remained unresolved. This separation prevents a chain of optimistic labels from substituting for a chain of evidence.

Add a trade-off sheet with environmental and service requirements. It should include absolute totals and meaningful rates, significant exclusions, burdens on different places or people, and any non-negotiable constraints. Use a sensitivity test to identify the assumptions that change the ranking. An uncertainty that never changes the decision may deserve a different research priority from a small uncertainty that flips the preferred option.

Finally, write a decision record. State the narrow conclusion, the evidence supporting it, the strongest limitation, the next justified action and the condition that would change the decision. Include a responsible reviewer and a date for reassessment where the real workflow requires one. The record should allow someone who did not run the model to understand what has and has not been established.

For Packet E, a completed record might read: “On stipulated batch measurements and factors, Willow has lower operational electricity-related emissions per accepted report. After the supplied manufacturing allocation, both workflows are 40 grams per accepted report. Willow's stipulated onsite consumptive water measure is higher. The estimates omit several lifecycle effects and do not establish downstream environmental benefit. We will compare equal service volumes and clarify allocation and water boundaries before making a broader claim.”

An AI assistant can help organise this packet, locate inconsistencies or propose checks. Give it a bounded instruction: use only the supplied figures for calculations, separate hypothetical assumptions from evidence, show units, and list missing information without inventing it. Then verify the arithmetic independently. The assistant is useful when it makes the evidence easier to inspect, not when it fills every blank with a plausible-sounding number.

Back to contents · Continue: 12. Independent exercises with complete answers

12. Independent exercises with complete answers

Exercise one is a classification check. A fictional test set has 300 images, including 60 reference-positive images. A system flags 75 images, of which 45 are reference-positive. Calculate true positives, false positives, false negatives, true negatives, precision, recall and accuracy. Then state one thing the result cannot tell us about biodiversity. Work from the counts before looking at the answer.

The answer is 45 true positives and 75 − 45 = 30 false positives. There are 60 − 45 = 15 false negatives. The 240 reference-negative images leave 240 − 30 = 210 true negatives. Precision is 45/75 = 60%; recall is 45/60 = 75%; accuracy is (45 + 210)/300 = 85%. The four counts sum to 300. The result does not establish a population trend or an ecological improvement; it measures classification against the supplied reference labels.

Exercise two concerns a missing denominator. A fictional report claims that a tool reduced energy use by 25%. The old workflow used 80 kWh for 400 accepted reports. The new one used 90 kWh for 600 accepted reports. Calculate both totals and energy per accepted report. Decide what part of the claim is supportable and rewrite it precisely.

The old rate is 80/400 = 0.2 kWh per accepted report; the new rate is 90/600 = 0.15 kWh. The rate falls by 0.05/0.2 = 25%. Total electricity rises by 10 kWh, or 10/80 = 12.5%, while accepted output rises by 50%. A precise sentence is: “Electricity per accepted report fell 25%, while total electricity rose 12.5% as accepted output increased from 400 to 600.” No net environmental benefit follows without a wider comparison.

Exercise three tests attribution. A fictional site uses 5,000 kWh before and 4,300 kWh after an intervention. A valid-for-this-exercise proportional comparison factor is stipulated to be 0.92. Calculate the observed reduction and the modelled intervention difference. Then say why a real evaluation would still need to examine the validity of the comparison factor.

The observed reduction is 700 kWh, or 14% of the 5,000 kWh baseline. The stipulated counterfactual is 5,000 × 0.92 = 4,600 kWh, so the modelled intervention difference is 300 kWh. That is 6% of baseline or about 6.52% of the counterfactual. The arithmetic is exact given the inputs, but a real attribution requires evidence that the factor represents the site's untreated change. Differences in activity, weather or maintenance could invalidate it.

Exercise four is a unit and boundary check. A fictional batch uses 20 kWh of IT electricity, with a facility-to-IT ratio of 1.25 and an invented emissions factor of 0.3 kg CO2e/kWh. It produces 100 accepted reports. Calculate facility electricity, electricity-related emissions and grams per accepted report. Explain why manufacturing must not be assumed included.

Facility electricity is 20 × 1.25 = 25 kWh. Electricity-related emissions are 25 × 0.3 = 7.5 kg CO2e. Per accepted report that is 7.5/100 = 0.075 kg, or 75 grams. Manufacturing was never supplied in the boundary or inputs, so it is excluded. A statement that these are total lifecycle emissions would be unsupported. If 20 kWh had already been measured at the facility boundary, the extra multiplication would be inappropriate.

Exercise five tests a multi-objective decision. Two fictional options have normalised climate and habitat scores. Aster has 0.8 and 0.3; Birch has 0.5 and 0.7. At equal weights, calculate both totals. Then calculate the climate weight at which they tie, using the complementary habitat weight. Does either option dominate the other on those two scores?

At equal weights, Aster scores 0.55 and Birch 0.60. With climate weight w, Aster is 0.3 + 0.5w and Birch is 0.7 − 0.2w. Equality gives 0.7w = 0.4, so the tie is at 4/7, about 57.14%. Neither dominates because Aster has the higher climate score and Birch the higher habitat score. These scores express the exercise's model; they do not determine who has legitimate authority to choose the weights.

Exercise six examines rebound. A fictional system performs 2,000 tasks at 0.05 kWh each. A new version uses 0.04 kWh per task and runs 2,400 tasks. Calculate old use, fixed-volume new use, actual new use and the fraction of expected savings taken back by expansion. What extra evidence is needed before calling expansion a causal rebound effect?

Old use is 100 kWh. At 2,000 tasks, the new version would use 80 kWh, creating expected savings of 20. At 2,400 tasks it uses 96 kWh, so actual savings are 4. Expansion takes back 16/20 = 80% of the fixed-volume savings. To make a causal claim, we need evidence that the efficiency change induced the extra activity, rather than assuming that all growth resulted from it.

Exercise seven is a writing test. Rewrite this fictional promotional claim: “Our AI achieved 90% accuracy and therefore restored the ecosystem while offsetting all its computing emissions.” Assume the only evidence available is a held-out image classification test with no ecological monitoring, lifecycle calculation or credit verification.

A defensible rewrite is: “The system achieved 90% accuracy on the specified held-out image test. The available evidence does not establish ecosystem restoration or compensation for computing-related emissions.” A stronger report would also provide the class balance, confusion matrix, relevant error metrics and sampling method. The right repair is to narrow the claim and request the missing evidence, not to add more confident adjectives.

Check your independence by completing these exercises without asking a model which answer sounds best. You should be able to identify the numerator, denominator, unit, boundary and evidential status of every result. If an answer differs, trace the mismatch to one of those elements. Repeating the same prompt until an assistant agrees is not an independent verification method.

Back to contents · Continue: 13. Common questions about SI and the environment

13. Common questions about SI and the environment

Can current AI help environmental work? It can support bounded tasks such as organising evidence, analysing authorised data, classifying observations and comparing modelled options. Usefulness must be demonstrated for the particular task and conditions. A capable assistant can make work faster while still needing expert checking, suitable instruments and legitimate decisions. Claims of broad environmental benefit require evidence beyond fluent outputs or isolated benchmarks.

Would hypothetical ASI automatically solve climate change? That conclusion does not follow from higher intelligence. Physical deployment, resource constraints, institutions, coordination, rights and competing goals would still matter. A hypothetical system might identify options that people currently miss, but their effects and implementation would need evaluation. Speculation about future capability should not be used to postpone justified present action or excuse unverified promises.

Is biodiversity just another carbon metric? No. Carbon-related accounting addresses particular climate quantities. Biodiversity includes several dimensions of living systems and their context. A project can improve one measure while damaging another. Even a useful habitat index has a scope and method; it should not be treated as a universal currency that makes different places or species interchangeable.

Does lower prediction error prove a better environmental outcome? No. It establishes a performance improvement on the relevant prediction test, assuming that the test is valid. The prediction still has to inform an effective, implemented and legitimate intervention. An accurate warning that never reaches the people who need it may fail to produce the intended benefit. The decision and delivery pathway must be evaluated as well.

Can a project be worthwhile if its computing uses electricity and water? Possibly, but the answer depends on the service, alternatives, outcomes and wider burdens. A footprint is part of the comparison, not a complete verdict by itself. Count it transparently and avoid treating an assumed downstream benefit as a verified cancellation. The same discipline applies to a project that claims no benefit: compare realistic options rather than slogans.

Should every uncertainty stop a decision? No. Decisions often have to be made with incomplete information. The appropriate response depends on consequences, reversibility, available alternatives and the value of further evidence. A low-risk desk exercise can proceed with labelled assumptions. An irreversible intervention affecting habitats or people needs a much stronger basis and appropriate authority. Stating uncertainty helps calibrate action rather than simply producing paralysis.

What should a student do when sources disagree? First check whether they describe the same quantity, place, time, reference period and accounting boundary. Then examine methods, uncertainty, source quality and whether one source is a scenario rather than an observation. Some disagreements remain substantive. Record them honestly instead of asking an AI system to manufacture a false consensus. A well-defined unresolved question can be a successful learning outcome.

What is the most useful single habit? Ask, “What exactly would have to be true for this conclusion to follow?” Then identify which of those conditions is measured, assumed or missing. This question works for a satellite map, a forecast, an emissions estimate and an optimisation result. It keeps the reader's judgement active while allowing AI to contribute useful analytical work.

Back to contents · Continue: 14. A final reader checkpoint

14. A final reader checkpoint

You have completed the guide when you can do more than repeat that AI has both benefits and risks. You should be able to take one environmental claim and separate observation from prediction, attribution, optimisation and authority. You should be able to name the comparison, recover the denominator, check the units and explain the largest limitation without either exaggerating it or hiding it.

Try the checkpoint on a fresh, public claim. Write its exact scope in one sentence. List the evidence that would support that scope, then mark the evidence actually supplied. Choose the smallest defensible conclusion. Finally, propose one next step that would materially improve the decision: a better comparator, an independent test, a local review, a missing footprint boundary or a clearer statement of objectives.

That is the role of environmental evidence literacy in a Super Intelligence world. More capable systems can help people inspect more information and explore more options. Better decisions require that the resulting claims remain checkable, that environmental burdens stay visible, and that people retain legitimate ways to choose and challenge what happens in their places.

Back to contents · Continue: Sources and further reading

Sources and further reading

  • IPCC, AR6 Synthesis Report: Climate Change 2023, with its Summary for Policymakers. Assessment context for mitigation, adaptation, risk and inclusive climate-resilient development: Read source
  • IPCC, AR6 Synthesis Report Summary for Policymakers, especially section C.6 on governance, participation and diverse knowledge: Read source
  • WMO, State of the Global Climate 2025, published 23 March 2026. Dated global observation context, not a local AI intervention evaluation: Read source
  • IPBES, Nexus Assessment, 2024. Interconnections among biodiversity, water, food, health and climate: Read source
  • IPBES, Business and Biodiversity Assessment, 2026. Methods for assessing impacts and dependencies: Read source
  • UNEP, Artificial Intelligence end-to-end: The Environmental Impact of the Full AI Lifecycle Needs to be Comprehensively Assessed, 2024: Read source

Continue learning

Return to What is Super Intelligence? for the definition and evidence map. For the wider research process, continue with scientific discovery and verified knowledge. For the separate infrastructure question, read how AI demand can finance energy.

Back to contents