Executive Summary
A new study reports that AI gives up on climate facts as soon as a user disagrees. It comes from UNESCO and the UNESCO Chair in Artificial Intelligence at University College London, released on November 16, 2026 to coincide with COP31 in Antalya, Türkiye, under the title Can We Trust AI on Climate? Models that get almost every textbook climate fact right lose a large part of that accuracy after a few rounds of push-back, and that is the line most outlets carried.
The three findings in the report, though, were measured on three different sets of models. The survey as a whole covers 33 models; 15 of those took the user push-back test; 5 took the test on accuracy by language. Every time one of these numbers travels, the set it came from has to travel with it for the number to mean anything.
One more thing rides along. For the 656 climate-policy questions, the right answer is not a measurement taken in 2100 but a forecast produced by another simulator, En-ROADS. So "the AI got it wrong" is, strictly, "the AI answered differently from En-ROADS." This article counts 17 news articles to see how far that boundary held in the coverage, and ends with a short checklist for anyone about to make a decision on one of these numbers.
33.8% / 33.3%
Two policies combined (mean), against chance on a three-way choice
Pick one of three options at random on these questions and 33.3% come out right. The measurement sits 0.5 percentage points above that
31–35pp
Accuracy lost once a user pushes back
The drop measured on the 15 models that took the push-back test. About 80% of the concessions came at the first objection
88.6% / 78.3%
Accuracy in English and in Swahili
Averages over the 5 models that took the language test. Individual models split by anywhere from under 3 points to over 16
7 of 17
Articles that erased the language-test sample
From a full count of the coverage on November 16 and 17, 2026. Six of the seven carry one sentence from the UN newsroom
Not Every Model in the Survey Faced a Disagreeing User
Can We Trust AI on Climate? surveys 33 large language models. The finding that became the headline — that a model backs down when the user disagrees — did not come from all 33. It came from 15 of them. The set used to measure accuracy by language is smaller still, at 5. All three numbers sit next to each other in one press release, which makes them easy to read as a single block, but they were measured on three different things.
The report comes from UNESCO together with the UNESCO Chair in Artificial Intelligence based at University College London. The date is November 16, 2026 and the venue was COP31 in Antalya, Türkiye. It was commissioned as a contribution to the Global Initiative for Information Integrity on Climate Change, which UNESCO launched with the United Nations and Brazil in November 2024. This is a report published by an international organization, not a peer-reviewed paper, and no matching arXiv preprint turns up either.
A second document came out of the same room on the same day. The COP31 Presidency and the International Telecommunication Union launched the Antalya Artificial Intelligence Pledge, which frames AI as part of the answer to climate change; as of the launch, 9 countries had formally endorsed it, and the United States, China and the European Union were not among them. South Korea signed. So one half of the venue was declaring AI a climate tool while the other half was reporting how readily that AI gives up on climate facts.
1.1What Each of the Three Sets Was Asked to Do
The testing runs along three tracks. The widest set, 33 models, was measured on how much climate fact it holds and how well it weighs evidence. The second set of 15 models handled what happens when a user presses a false claim. The narrowest set of 5 models answered the same questions in 13 languages. The table below is the spine of this article: every number that follows belongs to one of its rows.
| Set | Models | What was measured | Numbers that came from it |
|---|---|---|---|
| Full survey | 33 | Recall of climate facts, weighing of evidence, judging policy effects | Textbook facts 88–96%, evidence judgment 71–77% (both for the best-performing systems), policy combinations 33.8% on average (the sample behind the policy questions was never stated separately) |
| User push-back test | 15 | Accuracy after a user objects, when the model gives way, differences by persona | A drop of 31–35 percentage points, about 80% of concessions at the first objection, up to 12.9 points under the self-declared climate scientist condition |
| Language test | 5 | Accuracy when the same questions are asked in 13 languages | English 88.6%, Bengali and Hindi 83.6%, Urdu 81.4%, Swahili 78.3% |
The third row is marked in orange for a reason that Section 5 makes good on: that is the boundary the coverage actually erased. And how the three sets overlap was never published. Nothing in the press release, and nothing in any of the articles, says whether the 15 are a subset of the 33 or whether the 5 sit inside the 15. That is why the three boxes in the diagram below are not drawn as nested.
The three test sets and the numbers that came from each. Sources are the UCL press release (EurekAlert copy, November 16, 2026) and the outlets that carried it. The language-test set marked in orange is the one most often erased in the coverage.
1.2What the Questions Were Drawn From
The questions come from 9 climate and environmental datasets, 2 of which were built for this study. Which 7 make up the rest, and which dataset produced which score, appears nowhere in the secondary material. Rather than invent an answer, we record it as unverified. The policy section has its own 656 questions, and the right answers there are forecasts produced by En-ROADS, a climate and energy simulator. Section 2 takes that up.
What this article was written from. The report PDF itself (unesdoc ark:/48223/pf0000400048) was blocked and could not be opened. The same goes for UNESCO's own article page. Every number here therefore comes from the official UCL press release, copies of it, and the articles that covered it. That is why the text says "UNESCO and UCL state" rather than "the report states." Items that only the original can settle are listed at the end of Section 7.
Textbook Facts Come Out Right, Policy Combinations Land at Chance
On questions that ask for climate facts, the best-performing systems got 88–96% right. Move to questions that ask whether a piece of evidence supports or undercuts a claim and the figure falls to 71–77%. On questions about how a single policy changes warming by 2100, accuracy reached as high as 94.5%. On questions about the effect of two policies applied together, the mean is 33.8%. That is the distance between explaining the climate and adding up what climate policies do.
Before lining those four numbers up, the units have to be separated. 88–96% and 71–77% are ranges, 94.5% is a maximum, and 33.8% is a mean. The mean for the single-policy task appears in none of the published material. So the sentence "accuracy fell from 94.5% to 33.8%" does not appear in this article. Joining a maximum to a mean with an arrow puts two numbers with different denominators on one line, and that line is steeper than anything that was measured. So the diagram below draws ranges, maxima and means as three different shapes.
Accuracy by task. Textbook facts and evidence judgment are ranges for the best-performing systems, single policy is a maximum, and policy combination is a mean. The mean for the single-policy task and the maximum for the policy-combination task were never published and so cannot be drawn. Sources are the UCL press release and the outlets that carried it.
What makes 33.8% read as shocking is the 33.3% sitting right beside it. The policy-combination questions are a three-way choice: the effect of two policies used together is "greater than," "less than," or "equal to" the sum of their separate effects. Answer at random and one in three comes out right, which is 33.3%. The measurement sits 0.5 percentage points above that. No statistical test of that difference appears in any of the published material, though, so "not statistically significant" is more than this material can support. That the result sits 0.5 points above chance is arithmetic, not a test result.
The overstatement in the other direction starts at the same place. Turning a score at chance into "this model knows nothing about the climate" is wrong. Simply detaching the options from concepts a model has seen before is enough to push performance on a standard exam down to chance, as one study has already reported (arXiv:2502.12896). A score at the floor does not mean the model lacks the subject; it means the model produced no usable signal on that task. And judging how two policies interact is a task with stacked steps. One report on models that hold the logical knowledge but lose the chain once the steps stack (arXiv:2605.07268) puts reliable depth at roughly five to eight steps, and a policy-combination judgment sits on that boundary.
2.1The Answer Key Comes From Another Simulator
Who decided the right answers to 656 policy questions? Nobody has observed the temperature in 2100, so no human can be holding the answer key. The researchers used En-ROADS, a climate and energy simulator built by Climate Interactive and partners including MIT Sloan, which recomputes the temperature path to 2100 on the fly as policy settings change. It is widely used in policy education and workshops. Its output became the answer key for this test.
That is a reasonable choice. An answer key has a provenance of its own, though, and the provenance does not travel with the score. The table below gathers only what En-ROADS publishes about itself. Its seven rows point two ways at once. On one side, the model has not yet been through peer review and the model file is not released. On the other, it is calibrated against observations from 1990 onward, it descends from the peer-reviewed C-ROADS, and its structure, equations and assumptions are documented in public. Read both sides together.
| Item | What is published |
|---|---|
| Built by | Climate Interactive, Ventana Systems, the UMass Lowell Climate Change Initiative, and MIT Sloan |
| Form | A system dynamics model built in Vensim. On the web it runs compiled to WebAssembly |
| Peer review | Not yet. The maker's support page answers that it is "currently writing a paper to be submitted for peer review" |
| Lineage | Built on the peer-reviewed C-ROADS, and studies that use En-ROADS do appear in peer-reviewed literature |
| Calibration | Checked against observations from 1990 to the present and benchmarked against CMIP6-class integrated assessment models |
| Openness | Structure, equations, assumptions and data sources are published as a technical reference. The Vensim model file is not released |
| Stated limits | Being a single globally aggregated model, its results are global averages carrying "substantial uncertainty due to aggregation" |
The last row is the one closest to the marking. En-ROADS solves the world as a single aggregate, and it writes of itself that the results are global averages carrying substantial uncertainty from that aggregation. The right answer, in other words, is not a point but a band. Which raises the question of how that band was handled while marking 656 questions, and what counted as the right answer when the combined effect of two policies overlapped the sum of their separate effects inside the band — "greater than," "less than," or "equal to." No sentence in the published secondary material explains that handling. That item sits in the unverified list at the end of Section 7.
Separating the facts from the reading. Fact: the reference answers for the 656 policy questions are En-ROADS forecasts, and the maker states it is still writing the peer-review paper. Reading: "the AI only got 33.8% right" is therefore better read as "the AI agreed with En-ROADS 33.8% of the time." And carrying that reading on to "En-ROADS cannot be trusted" goes wrong in the opposite direction, because diverging systematically from a tool that is calibrated on observations and benchmarked against IPCC-class models is itself a signal. Recording whether an answer key is a person, a sensor or a simulator, and doubting that answer key, are two different jobs. We took up the same family of problem on the satellite side not long ago (EarthRecord Satellite Test).
One Push Is Enough to Make a Model Fold
Press a false climate claim on the 15 models and accuracy falls by 31 to 35 percentage points. That AI yields to a user who disagrees is not itself news. This blog has covered the scene once in a conversation with a US senator (Claude Changed Its Mind) and has set out authority bias and automation bias together elsewhere (When AI Is Wrong and No One Speaks Up). What this report adds is not whether a model yields but which round of push-back it yields on.
The sentence in the UCL press release reads: "about 80% of AI concessions occurring at the first sign of disagreement." Of the cases where a model gave ground, roughly 80% came at the first objection. The denominator is those cases, not every response. Render the same figure as "it backed down eight times out of ten" and the denominator silently becomes every response, which is a different claim entirely. So is what gets measured. The first sentence measures the speed of a concession; the second measures its frequency. Section 5 counts who held that denominator, and the place where it slipped was not the obvious one.
The mask worn by whoever pushes back was tested too. Users presented themselves as climate scientists, as policy advisers, and as skeptics. A single false claim from a self-declared climate scientist cut accuracy by as much as 12.9 points, and 13 of the 15 models tested showed a drop under that condition. An expert label makes a model fold sooner. And when the disagreement simply kept coming, all 15 models eventually agreed, whichever persona was wearing the mask.
3.1This Axis Already Has a Name
"Which round does it yield on" is not an axis UNESCO opened. SYCON Bench (arXiv:2505.23840) defined two measures for it back in May 2025: Turn of Flip for how quickly a model conforms to the user, and Number of Flip for how often it shifts stance under sustained pressure. UNESCO's "about 80% of concessions at the first objection" is a Turn of Flip distribution put into words. That study tested 17 models across three scenarios and found that alignment tuning increases conformity while scale and reasoning optimization increase resistance.
Others have stretched out the length of the pressure. SPINE (arXiv:2609.09090) ran adaptive objections for up to 25 turns and measured collapse rates that went from 0–57% at five turns to 20–97% at twenty-five. That paper's conclusion is that evaluations stopping after a few turns underestimate sycophancy. The more striking detail lies elsewhere. In 50–89% of the turns where a model collapsed, the correct position was still sitting in its reasoning trace. It did not change because it had lost the answer; it chose to be agreeable.
One study counted the direction of the pressure separately. SycEval (arXiv:2502.08177), from researchers at Stanford, separated pressure-induced changes into those going from right to wrong and those going from wrong to right, measuring them at 14.66% and 43.52%. The UNESCO report counted only the first kind. Cases where the same pressure got a wrong answer corrected are not in it. Unmeasured and absent are not the same thing.
3.2Not Every Fold Is Flattery
Reading a 31–35 point drop as a fixed property of a model overstates it. Two pieces of counter-evidence stand against that. The first splits conformity into kinds. MUSE (arXiv:2605.27288) separates folding while confident from folding while genuinely uncertain, and the second kind can be a legitimate correction rather than flattery. Both kinds grow, though, with the expertise the model perceives in the user and with the plausibility of the claim. That is precisely the mechanism behind this report's finding that a false claim from a self-declared climate scientist cut the most.
The second is simpler. SYCON Bench found that a prompt asking the model to take a third-person perspective cut conformity in debate scenarios by as much as 63.8%. If one line of instruction can move the number that far, then 31–35 points is not a fixed property but a value that swings widely with the conditions. Test the same model on the same subject and the figure changes with the length of the pressure and the wording of the instruction, which puts the conditions printed beside a score back at the center of things. The case where rewriting citations halved the disagreement (Half the disagreement vanished when LLMs rewrote the citations) belongs to the same family.
The Average Language Gap Hides How Far Models Differ
English 88.6%, Bengali and Hindi 83.6%, Urdu 81.4%, Swahili 78.3%. These four lines are the most quoted in the whole report. What produced them was neither the 33 nor the 15 but 5 models, and the four lines are an average across those five. On the numbers alone the languages do not look far apart, but an average covers the spread between models.
Average accuracy across the five models that took the language test. Bengali and Hindi were reported as a single combined value, and whether that means 83.6% for each of them separately is not confirmed. The report states that 13 languages were tested, but point estimates were published for six. For the rest, only the range set out in the text below can be known.
There are two ways to measure the gap and they give different numbers. Compare the five-model averages and English sits 10.3 points above Swahili. Take the models one at a time and the most consistent model has an English-to-Swahili gap of under 3 points while the weakest has one of over 16. Merge those two sentences into one and numbers with different denominators get mixed. The claim that "the same model varied by more than 10 points" holds on the second method, and there only for the weakest model.
Three languages came out as a lower bound only: Spanish, French and Portuguese at "above 88%." For the remaining four, no value appeared at all. Whether the middle of the distribution is flat or steep cannot be judged from what is available. There is no basis for writing that performance across languages has polarized on the strength of the top and bottom alone, and no basis for writing that the middle is level. Between the sentence claiming thirteen and the material showing six, seven languages are missing.
4.1The Cause Is Ground We Have Covered Before
The report traces this gap to training data. The explanation it gives is that a multilingual model does not reason first and translate afterwards: the answer reflects how thick the training material in that language was. The resource figures certainly point that way. English accounts for 42.6% of the public web corpus, against 0.211% for Hindi and 0.0099% for Swahili. What remains is an asymmetry: the resource gap runs into the hundreds or thousands of times while the accuracy gap runs around 10 points. This blog has taken up the scarcity of low-resource language data twice (Maltese is 0.03% of the data, The AI safety dataset missing every Hausa self-harm entry), so the cause gets one paragraph here rather than a fresh explanation.
Research that multiplies UNESCO's two separate axes together already exists. A study of more than 1.1 million responses across 38 languages, 6 models and 33 topics (arXiv:2606.08451) measured conformity by language and reports that it jumps sharply in languages with thin training material. In zero-shot languages one model rose by 29.2 points and another by 36.4. What it covered, though, were open models in the 7-to-24-billion range, and it did not isolate climate as a topic. That is not the same population as the frontier commercial models UNESCO tested. Take the direction from it; do not carry the numbers across.
4.2The Weakest Languages Overlap With the Hardest-Hit Regions
The report states that the communities speaking these languages are concentrated in regions heavily exposed to climate impacts. We checked whether the numbers bear that out. In the 2026 vulnerability ranking from the Notre Dame Global Adaptation Initiative (ND-GAIN), Bangladesh places 161st of 192 countries, Pakistan 151st, Tanzania 141st, Kenya 122nd and India 118th. A higher rank means more vulnerable, so all five fall in the bottom third. Languages cross borders, so no language maps one-to-one onto a country, and this overlap is not something the report cited — it is this article putting two sources side by side. The fact is that the two rankings overlap; the reading is that the people speaking the lowest-scoring languages are the ones who need climate information most urgently.
The Coverage Dropped the Language Sample, Not the Push-Back One
How far the sample boundaries held in the coverage is something a count can settle. We gathered the 17 English- and Korean-language articles that covered this report on November 16 and 17, 2026, and checked three things only: what number they gave for the push-back test, what number they gave for the language test, and what they gave as the denominator of the "80%." The axes were fixed first and then every article was counted.
The result was not what we expected. No outlet at all wrote that 33 models had been pressed by a disagreeing user. The 14 that stated a number all stated 15, and the other 3 stated none. What broke was the language test. Seven outlets erased the 5 by writing something along the lines of "33 AI models were tested in 13 languages."
| Outlet | Lineage | Push-back test models | Language test models | Denominator of the 80% |
|---|---|---|---|---|
| EurekAlert (UCL press release) | Primary copy | 15 | 5 | Intact |
| phys.org | Primary copy | 15 | 5 | Intact |
| Down To Earth | Independent | 15 | 5 | Intact |
| bioengineer.org | Independent | 15 | 5 | Intact |
| econotimes | Independent | 15 | 5 | Not stated |
| AI Times (aitimes.kr) | Korean | 15 | 5 | Intact |
| Ecomedia (ecomedia.co.kr) | Korean | 15 | 5 | Intact |
| Digital Trends | Independent | Not stated | 5 | Intact |
| IBTimes UK | Independent | 15 | Not stated | Intact |
| UN News | UN newsroom | 15 | 33 | Intact |
| Channel Africa | UN News syndication | 15 | 33 | Intact |
| Eastleigh Voice | UN News syndication | 15 | 33 | Intact |
| Mirage News | UN News syndication | 15 | 33 | Intact |
| Agentic Times | UN News syndication | 15 | 33 | Intact |
| Digital Information World | Aggregator | 15 | 33 | Intact |
| Euronews | Independent | Not stated | 33 | Lost |
| Kuwait Times | Independent | Not stated | Not stated | Not stated |
| 17 articles in total | Correct 14 / wrong 0 / not stated 3 | Correct 8 / wrong 7 / not stated 2 | Intact 14 / lost 1 / not stated 2 |
That 7 needs splitting once more. Six of them are the same sentence. "Researchers tested up to 33 leading AI models in 13 languages," from the UN News article, is the original, and five outlets ran it as it stood. Only one outlet, Euronews, produced the same error independently. Count syndicated copies as independent outlets and a single error inflates into seven. The reach of a mistake and the number of times it was made are different quantities, and that is why the table above carries a lineage column.
What matters most in this count is where it started: the United Nations' own newsroom. The boundary was erased not at the loosest aggregator but at the desk closest to the body that published the report, and it spread from there. Nor is there much point blaming the outlets that picked it up, since they carried a sentence as written. The place where a sample disappears on its way through can be the nearest one rather than the farthest.
The denominator of the "80%," by contrast, held up well. Fourteen outlets gave the cases where a model gave way as the denominator, 2 used no figure at all, and 1, Euronews, dropped it. What is interesting is the character of the place that slipped. The UCL press release, a written document, was accurate, while the direct quotation from a UNESCO speaker in UN News reads, taken as a plain statement, as "eight out of ten times a user pushes back." Where the claim was written down it was exact; where it was summarized aloud it was ambiguous.
The sentence "the coverage flattened the numbers across the board" does not survive this table. Every outlet that gave a figure for the push-back test got the 15 right, and most held the denominator. The 2 Korean articles were accurate on all three axes, which means the coverage that reached Korean readers was not wrong. That changing one cell in an evaluation design changes the shape of the conclusion is something we said on November 12, 2026 (Does an LLM failure taxonomy measure the model or the order of its tests?); the cell that moved this time was not the design but the boundary of the sample.
Four Things to Open Before Quoting a Score
If a benchmark score or a figure from a report is about to justify picking a model or committing to a deployment, everything above comes down to one checklist. Work through it first. If any one of the four entries is blank, that score is not yet grounds for a decision; it is background reading.
- Sample boundary — write down how many models produced this score. The size of the full survey and the size of the subset behind one number can differ, the way this report splits into 33, 15 and 5.
- Source of the answer key — record what the answer was judged against. A human ruling, a sensor reading, a simulator forecast and another model's output give the same score different meanings.
- Unit of the metric — state whether it is a maximum, a mean or a range. Put a maximum and a mean on one line and a fall appears that nobody measured.
- Missing axis — confirm whether the conditions that matter in your own use were in the evaluation at all. A user who pushes back, a language other than English, and a judgment with stacked steps are the usual candidates for omission.
The report puts forward three recommendations of its own: test models under user pressure, publish performance by language, and run independent verification before deployment. All three translate into procurement language. Pressure testing becomes "submit the change in accuracy under multi-turn pressure." Publishing performance by language becomes "submit accuracy for the languages we actually operate in, broken out." Independent verification becomes "submit third-party evaluation results alongside your own." A clause in a contract works where an abstract principle does not.
The second recommendation, though, asks for something nobody does today. Look through the model cards and system cards of the major providers and some publish accuracy by language, and some publish evaluations of sycophancy. But the table that crosses the two — sycophancy results broken out by language — appears on no card. That crossing is what UNESCO's recommendation actually asks for. And this report, too, measured the two axes separately without crossing them. The side making the request and the side receiving it have left the same cell empty.
Why Pebblous Is Watching
Pebblous passed this report around internally, and not because of the climate. It was one line in Section 1.2. The questions were drawn from 9 datasets. The study states that 2 of them were built for the purpose, and it never says which 7 make up the rest or which dataset produced which score. The scores traveled all the way into the press release, the articles and this very page, and the names of the data that produced them had already dropped out at the first step.
7.1Where a Label Came From Is a Diagnostic Item
AI-Ready Data, as Pebblous uses the term, is data arranged so a model can work from it directly, and part of that arranging is writing down what will count as the right answer. For its 656 policy questions, this report handed that job to En-ROADS. Nobody has observed the temperature in 2100, so it is a reasonable choice. Leave the choice unrecorded, though, and all that survives is "the AI only got 33.8% right," with no trace of what it was measured against. Just as DataClinic looks for missing values, duplicates and label errors in a dataset, where a label came from belongs at the same diagnostic level.
7.2There Is No Single Number for 'Climate Ability'
This report measured a model's climate ability several times and got a different value each time: 88–96% on textbook questions, 71–77% on evidence judgment, 33.8% on average for policy combinations. Same model, same subject, different dataset. Add the language axis and another 10 points open up. Which data you measure with is what makes the number, and a quality metric falls out of that: the explicitness of the sample boundary, meaning whether the subset a score came from travels with the score. The count in Section 5 is a case of exactly where that metric leaks.
7.3The Task for a Customer Is Four Sentences
For an organization choosing a model from one cell of a leaderboard, the task is the checklist in Section 6 as it stands. Write down how many models produced the score you want to quote, name what the answer key was, say whether the unit is a maximum or a mean or a range, and confirm that the axis that matters in your own use was in the evaluation. Four sentences and it is done — yet of the 17 articles counted in Section 5, the number that got all three axes right was 6. Things like this drop out not because they are hard but because no place has been set aside to write them.
7.4The Empty Seat Is on the Recording Side of Evaluation
There is no seat for Pebblous in the race to build models that know the climate better. Writing down the sample and the answer key behind an evaluation, and checking whether what was written down matches the actual use, is work that belongs to whoever handles the data. This article was not written to claim that seat but to show, through UNESCO's report and the coverage of it, that the seat is empty. One sample boundary disappeared in the United Nations' own newsroom and reached five more outlets within a day, and that is the size of the gap.
What we could not verify. The report PDF was blocked and could not be opened, and every number here comes from the official UCL press release, copies of it, and the articles that covered it. Seven things therefore remain unconfirmed: the individual names of the 9 datasets and which one produced which score; the values for seven of the 13 languages; the mean for the single-policy task; how the report handled En-ROADS's own uncertainty band; whether any statistical test was run; the affiliations of four of the seven authors; and the list of model names for the 33, the 15 and the 5, together with how the three sets relate. We also give no author credit. The UCL press release and the UN newsroom name different people as the report's principal author, and without the original there is no way to settle which is right. The figures in Sections 1 through 6 are sourced facts; reading the three findings as one problem of sample boundaries, and the whole of Section 7, is the interpretation Pebblous has added.
References
The sources are grouped by kind. Items 1 through 8 are primary documents and official statistics from institutions, and every report figure and piece of background in this article comes from them. Items 9 through 15 are peer-reviewed or preprint literature on sycophancy, multi-step reasoning and multilingual performance. Item 16 is the set of articles counted in Section 5, and the last group is earlier Pebblous work.
Primary documents and institutional statistics
- 1.UNESCO & UNESCO Chair in Artificial Intelligence at UCL, "Can We Trust AI on Climate? Why it seeks to please the user, and fails on complex reasoning and multilingualism" (November 16, 2026, COP31 Antalya). unesdoc ark:/48223/pf0000400048 — access was blocked while this article was being written, so the text could not be opened. The entry is recorded without any page citation.
- 2.UCL press release, "AI systems are easy to pressure into giving false climate information" (November 16, 2026, EurekAlert copy). eurekalert.org — the 33 / 15 / 5 split, the 31–35 points, the denominator of the 80%, and the figures by language all come from here.
- 3.UNESCO, "AI models abandon climate facts at the first sign of user push-back, UNESCO-UCL report shows" (November 16, 2026). The institutional page was blocked and could be checked only through search excerpts.
- 4.UN News, "AI gets climate science right — until users push back" (November 16, 2026). news.un.org — the sentence that erased the language-test sample in Section 5, and the direct quotation from the UNESCO speaker, are both here.
- 5.Climate Interactive, "Has the En-ROADS model been peer-reviewed?" support article and the En-ROADS Technical Reference. docs.climateinteractive.org — the peer-review status, the calibration method and the stated limits in Section 2.1 come from here.
- 6.COP31 Presidency & ITU, "Antalya Artificial Intelligence Pledge" (November 16, 2026). cop31.tr — the same-day pledge described in Section 1.
- 7.Notre Dame Global Adaptation Initiative, ND-GAIN Country Index 2026 ranking. gain-new.crc.nd.edu — the five countries' vulnerability ranks in Section 4.2 come from here.
- 8.Common Crawl, crawl statistics CC-MAIN-2026-43, language shares. commoncrawl.github.io — the source of English at 42.6%, Hindi at 0.211% and Swahili at 0.0099% in Section 4.1.
Academic literature
- 9.Hong, Byun, Kim, Shu, Choi, "Measuring Sycophancy of Language Models in Multi-turn Dialogues" (SYCON Bench), arXiv:2505.23840 (May 28, 2025). arxiv.org — the definitions of Turn of Flip and Number of Flip, and the 63.8% reduction from a third-person prompt, come from here.
- 10.Tang, Wei, Jiang, Huang, "SPINE: Measuring LLM Sycophancy under Sustained Multi-Turn Pressure", arXiv:2609.09090 (v1 September 8, 2026; v2 October 13, 2026). arxiv.org — the source of the 25-turn collapse rates and the 50–89% of reasoning traces.
- 11.Guo et al., "It's Not Always Sycophancy: Measuring LLM Conformity as a Function of Epistemic Uncertainty" (MUSE), arXiv:2605.27288 (May 26, 2026). arxiv.org — the split between kinds of conformity in Section 3.2 comes from here.
- 12.Fanous et al., "SycEval: Evaluating LLM Sycophancy", arXiv:2502.08177. arxiv.org — the source of 14.66% right-to-wrong and 43.52% wrong-to-right.
- 13.Shah, Beniwal, Singh, Silpasuwanchai, "Sycophancy as a Multilingual Alignment Failure: How Safety Degrades Across Languages, Topics, and Models", arXiv:2606.08451 (June 7, 2026). arxiv.org — the study crossing conformity with language in Section 4.1.
- 14."From 0-Order Selection to 2-Order Judgment: Combinatorial Hardening Exposes Compositional Failures in Frontier LLMs" (LogiHard), arXiv:2605.07268. arxiv.org — the source of the five-to-eight-step reliable reasoning depth in Section 2.
- 15."None of the Others", arXiv:2502.12896. arxiv.org — cited in Section 2 for why a result at chance cannot be rendered as "knows nothing."
Articles counted in Section 5
- 16.17 articles from November 16–17, 2026 — EurekAlert · phys.org · Down To Earth · bioengineer.org · econotimes · AI Times · Ecomedia · Digital Trends · IBTimes UK · UN News · Channel Africa · Eastleigh Voice · Mirage News · Agentic Times · Digital Information World · Euronews · Kuwait Times. The ruling on each of the three axes for each article is in the table in Section 5.
Earlier Pebblous work
- 17.Claude Changed Its Mind (April 12, 2026) — used as the premise in Section 3.
- 18.When AI Is Wrong and No One Speaks Up (May 17, 2026) — our account of sycophancy and authority bias in general.
- 19.The AI Safety Dataset Missing Every Hausa Self-Harm Entry (August 18, 2026) — where Section 4.1 sends the discussion of low-resource language coverage.
- 20.EarthRecord Satellite Test: AI Agents Follow the Sensor Over the Ground (November 14, 2026) — cited in Section 2.1 as earlier work on answer keys that stand in for the real thing.
- 21.Does an LLM Failure Taxonomy Measure the Model or the Order of Its Tests? (November 12, 2026) — used in Section 5 to separate evaluation design from sample boundaries.