Local News

AI’s Climate Science Failures Spark Warning for Vulnerable Caribbean Nations

17 November 2026
This content originally appeared on Trinidad Guardian.
Promote your business with NAN

“A chat­bot can be a bril­liant as­sis­tant. It can­not be your cli­mate ad­vi­sor.”

That stark warn­ing from UN­ESCO comes as new re­search re­veals that ar­ti­fi­cial in­tel­li­gence sys­tems can be per­suad­ed to aban­don es­tab­lished cli­mate sci­ence sim­ply be­cause users dis­agree with them.

The find­ings, un­veiled at the COP31 cli­mate con­fer­ence in An­talya, Türkiye, raise se­ri­ous con­cerns for Caribbean na­tions con­fronting ris­ing sea lev­els, de­struc­tive hur­ri­canes and mount­ing cli­mate adap­ta­tion costs.

One of the study’s most trou­bling find­ings was that rough­ly eight in ten in­stances of AI mod­els aban­don­ing sci­en­tif­i­cal­ly cor­rect an­swers oc­curred af­ter users first chal­lenged them.

The warn­ing came from Guil­herme Canela, UN­ESCO’s Di­rec­tor for Dig­i­tal Poli­cies and In­clu­sion, dur­ing the launch of the re­port, Can We Trust AI on Cli­mate In­for­ma­tion?, pro­duced in col­lab­o­ra­tion with Uni­ver­si­ty Col­lege Lon­don (UCL).

AI aban­dons facts un­der pres­sure

Re­searchers test­ed 15 AI mod­els and found that even mild ob­jec­tions could prompt sys­tems to re­verse sci­en­tif­i­cal­ly cor­rect an­swers, in­clud­ing es­tab­lished find­ings about car­bon diox­ide’s role in glob­al warm­ing.

The be­hav­iour, known as “AI syco­phancy”, oc­curs when sys­tems pri­ori­tise agree­ing with users over pro­vid­ing ac­cu­rate in­for­ma­tion.

“It is con­cern­ing that mod­els have a de­sign flaw … they are in­cen­tivized to agree with their user and de­ny these facts so read­i­ly,” said UN­ESCO lead au­thor Leona Ver­dadero.

Im­pli­ca­tions for Caribbean cli­mate pol­i­cy

The re­search al­so ex­posed weak­ness­es in AI’s abil­i­ty to analyse com­plex cli­mate poli­cies.

While the strongest mod­els achieved up to 94.5 per cent ac­cu­ra­cy when as­sess­ing in­di­vid­ual pol­i­cy mea­sures, av­er­age ac­cu­ra­cy plunged to 33.8 per cent when eval­u­at­ing how two poli­cies in­ter­act, bare­ly above ran­dom guess­ing.

That find­ing has im­pli­ca­tions for Caribbean gov­ern­ments weigh­ing in­ter­con­nect­ed de­ci­sions on re­new­able en­er­gy, emis­sions re­duc­tion and cli­mate re­silience.

UCL re­searcher Dr Maria Perez-Or­tiz warned that AI’s abil­i­ty to ex­plain sci­ence can cre­ate mis­placed con­fi­dence in its ca­pac­i­ty to guide de­ci­sions.

“Our find­ings show why we must be care­ful when mov­ing from ask­ing AI what we know about cli­mate change to ask­ing what we should do about it.”

Lan­guage di­vide rais­es con­cerns

Re­searchers al­so found that cli­mate in­for­ma­tion ac­cu­ra­cy var­ied sig­nif­i­cant­ly across lan­guages, av­er­ag­ing 88.6 per cent in Eng­lish but falling to 78.3 per cent in Swahili.

Al­though Caribbean Cre­ole lan­guages were not test­ed, the find­ings raise ques­tions about eq­ui­table ac­cess to re­li­able AI-gen­er­at­ed in­for­ma­tion across the re­gion’s mul­ti­lin­gual com­mu­ni­ties.

UN­ESCO is urg­ing de­vel­op­ers to im­prove mul­ti­lin­gual per­for­mance, strength­en ac­count­abil­i­ty and sub­ject AI sys­tems to in­de­pen­dent test­ing.

For Caribbean gov­ern­ments, jour­nal­ists and ed­u­ca­tors in­creas­ing­ly turn­ing to ar­ti­fi­cial in­tel­li­gence, the find­ings re­in­force the im­por­tance of ver­i­fy­ing AI-gen­er­at­ed in­for­ma­tion against cred­i­ble sci­en­tif­ic sources.