{"schema_version":"0.10.0","id":"LR-2026-0278","record_version":2,"title":"LLM agents in chemistry lab-reasoning study ignore experimental evidence and assert unsupported claims","verification":{"human_verified":false,"method":"automated_gate","auto_verified":true,"score":45.8,"score_breakdown":{"source_authority":12.1,"independent_corroboration":7.0,"implicated_party_confirmed":0.0,"citation_integrity":14.0,"specificity":0.0,"completeness":5.44,"unhedged_language":5.0},"checks":{"schema_valid":true,"every_field_cited":true,"quotes_verbatim":true,"source_fetched":true,"distinct_incident":true,"dates_plausible":true,"sources_agree":true,"source_describes_incident":true,"severe_backed_by_source":true,"record_integrity":true,"eligible":true,"about_this_case":true},"gate_version":"gate-2026-10-18.7","cross_check_model":null,"cross_check_agreement":null,"duplicate_of":null,"verified_by":null,"verified_at":"2026-10-30","extraction_run_id":null,"evidence_level":{"level":"preliminary","reason":"Publisher not yet named · subject to change","original_document":null,"reports_total":1,"reports_independent":0,"reports_derivative":0,"developing":true,"disputed":false,"rules_version":"evidence-2026-10-29.1","computed_at":"2026-10-30","history":[{"date":"2026-10-30","from_level":null,"to_level":"preliminary","reason":"Publisher not yet named · subject to change"}]},"sources_hash":"f8e13e7a80ab585cfcf215e1804fe9e8e943dc4556d53ad7285c74d98d5d1543"},"system":{"model_name":"ChatGPT","model_version":null,"developer":"unknown","harness":"AI agent","tools_granted":["digital tools","physical lab equipment"],"permissions_scope":null,"autonomy_level":"unknown","guardrails_present":[],"deployment_context":"research","tier":"provider","lifecycle_stage":"evaluation","integration_type":"agentic_with_tools","intended_use":"laboratory reasoning tasks, such as figuring out what was inside a mystery chemical solution","safeguard_state":null,"production_intended":null},"event":{"date_occurred":null,"date_disclosed":"2026-10-29","incident_class":"operational_error","failure_class":"silent_wrong_state","agent_involved":true,"secondary_failure_classes":[],"trigger":"unknown","duration":null,"detection_method":"unknown","summary":"Researchers gave three LLMs, including the one behind free ChatGPT, AI-agent tooling to run chemistry experiments and identify a mystery solution.","tags":[],"date_detected":null,"near_miss":null,"root_cause":"data","root_cause_detail":"The article attributes the behaviour to training: the LLM learns to reproduce the most likely or common information, so when an experiment's result is unlikely or unusual the AI ignores it and bases conclusions on past data.","failure_class_detail":"The agents ignored experimental evidence and made claims without supporting evidence; the article does not say they concealed this.","responsible_party":null,"adversarial_actor":null,"input_provenance":null,"time_to_contain":null,"operations_paused":null,"remediation":null,"precursor_events":null},"impact":{"direct_loss_usd":null,"remediation_hours":null,"records_exposed":null,"third_parties_affected_count":null,"severity":"minor","severity_because":"minor: The failures occurred inside a research evaluation of laboratory reasoning tasks, with no third party or production system affected.","harm_class":["operational"],"affected_description":"laboratory reasoning tasks, including a mystery chemical solution","protected_class_involved":null,"data_categories_exposed":[],"loss_range_usd":null,"reversibility":null,"downstream_exposure_count":null},"evidence":{"sources":[{"url":"https://www.snexplores.org/article/5-quirks-how-ai-thinks","publisher":"Science News Explores","date":"2026-10-29","type":"press","title":"These 5 quirks reveal how AI doesn’t think like we do","archived_url":null,"attachment":"first_report"}],"per_field_confidence":{"system.model_name":"D","system.harness":"C","system.tools_granted":"C","system.tier":"D","system.lifecycle_stage":"D","system.integration_type":"C","system.intended_use":"C","system.developer":"D","system.deployment_context":"D","system.autonomy_level":"D","event.failure_class":"D","event.root_cause":"D","event.agent_involved":"C","event.incident_class":"D","event.summary":"D","event.date_disclosed":"A","event.trigger":"D","event.detection_method":"D","impact.harm_class":"D","impact.affected_description":"D","impact.severity":"D","context.industry":"D","context.denominator_available":"D","context.country":"D","event.root_cause_detail":"D","event.failure_class_detail":"D","impact.severity_because":"D"},"field_citations":{"system.model_name":[{"source_index":0,"quote":"The researchers outfitted three different LLMs (including the one behind the free version of ChatGPT) with two types of AI-agent Iron Man suits."}],"system.developer":[{"source_index":0,"quote":"The researchers outfitted three different LLMs (including the one behind the free version of ChatGPT) with two types of AI-agent Iron Man suits."}],"system.harness":[{"source_index":0,"quote":"An AI-agent suit gives an LLM access to digital tools or physical lab equipment so it can take actions."}],"system.tools_granted":[{"source_index":0,"quote":"An AI-agent suit gives an LLM access to digital tools or physical lab equipment so it can take actions."}],"system.deployment_context":[{"source_index":0,"quote":"In one recent study, researchers showed that ignoring new evidence is actually the norm for LLMs."}],"system.lifecycle_stage":[{"source_index":0,"quote":"These systems then tackled laboratory reasoning tasks."}],"system.integration_type":[{"source_index":0,"quote":"An AI-agent suit gives an LLM access to digital tools or physical lab equipment so it can take actions."}],"system.intended_use":[{"source_index":0,"quote":"In one, they had to figure out what was inside a mystery chemical solution."}],"event.incident_class":[{"source_index":0,"quote":"In more than two of every three tasks, the AI agents ignored evidence at least once."}],"event.failure_class":[{"source_index":0,"quote":"And they made claims without any supporting evidence in more than half of the tasks."}],"event.failure_class_detail":[{"source_index":0,"quote":"In more than two of every three tasks, the AI agents ignored evidence at least once."}],"event.root_cause":[{"source_index":0,"quote":"Following its training, the AI often will ignore that result and instead base its conclusions on past data."}],"event.root_cause_detail":[{"source_index":0,"quote":"Following its training, the AI often will ignore that result and instead base its conclusions on past data."}],"event.agent_involved":[{"source_index":0,"quote":"These systems then tackled laboratory reasoning tasks."}],"event.summary":[{"source_index":0,"quote":"The researchers outfitted three different LLMs (including the one behind the free version of ChatGPT) with two types of AI-agent Iron Man suits."}],"impact.severity_because":[{"source_index":0,"quote":"These systems then tackled laboratory reasoning tasks."}],"impact.harm_class":[{"source_index":0,"quote":"And they made claims without any supporting evidence in more than half of the tasks."}],"impact.affected_description":[{"source_index":0,"quote":"In one, they had to figure out what was inside a mystery chemical solution."}],"context.industry":[{"source_index":0,"quote":"Do you have a science question? We can help!"}],"context.country":[{"source_index":0,"quote":"He’s a materials scientist at the Indian Institute of Technology Delhi in India."}],"impact.severity":[{"source_index":0,"quote":"These systems then tackled laboratory reasoning tasks."}],"event.date_disclosed":[{"source_index":0,"quote":"published: 2026-10-29"}]},"notes_on_gaps":"No source states event.date_occurred, event.date_disclosed, event.trigger, event.detection_method, event.duration, event.time_to_contain, event.date_detected, event.responsible_party, event.adversarial_actor, event.remediation, event.operations_paused, event.near_miss, system.model_version, system.permissions_scope, system.autonomy_level, system.guardrails_present, system.safeguard_state, system.production_intended, impact.direct_loss_usd, impact.records_exposed, impact.third_parties_affected_count, impact.reversibility, context.org_size_band, context.denominator, context.systemic_exposure, context.regulatory, context.coverage, or context.incident_evidence. system.developer is 'unknown' because the article names ChatGPT, Gemini, Grok and Claude but not the training organizations. event.trigger is omitted because no source names a cause of the behaviour beyond training. event.detection_method is omitted because no source says who first noticed. event.failure_class is graded D: the article describes a reasoning failure, not concealment, so silent_wrong_state is the closest mechanism. event.incident_class operational_error is graded D on the same basis. context.industry 'Educational Services' is graded D and rests only on the publisher's reader-question boilerplate, not on the operator. context.country is 'unknown' because the only country named is a quoted researcher's institution, not the operator's. impact.severity minor is graded D: the study is an evaluation with no third party touched, but no source states a loss figure or cleanup. system.tier: provider by rule (deployment_context research). [aggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-10-30] system.model_name: classified from described behaviour. system.lifecycle_stage: evaluation. system.developer: no source names the developer. system.deployment_context: a test or red-team run. system.autonomy_level: no source states the arrangement. event.incident_class: a source states data or a system was left in a wrong state, deleted, corrupted or acted on unintentionally, and names no adversary and no goal of the agent's own. event.summary: classified from described behaviour. event.trigger: no source states a cause. event.detection_method: no source says who first noticed. impact.harm_class: classified from described behaviour. impact.affected_description: classified from described behaviour. context.industry: classified from described behaviour. context.denominator_available: no denominator is available for this incident. context.country: no source states the operator's location."},"context":{"industry":"Educational Services","org_size_band":null,"country":"unknown","denominator_available":false,"denominator":null,"related_incident_ids":[],"systemic_exposure":null,"shared_model_id":null,"upstream_incident_id":null,"incident_evidence":null,"regulatory":null,"coverage":null},"narrative":{"steps":[{"step":"doing","text":"Researchers gave three LLMs, including the one behind free ChatGPT, AI-agent tooling with access to digital tools or lab equipment so they could run chemistry experiments and identify a mystery solution.","quote":"An AI-agent suit gives an LLM access to digital tools or physical lab equipment so it can take actions.","source_index":0,"grade":"C"},{"step":"went_wrong","text":"The agents ignored experimental evidence and made claims without supporting evidence, refusing to change a hypothesis even when evidence showed the line of investigation was wrong.","quote":"Even when you have clear evidence that shows that a particular line of investigation is not correct, [the AI] refuses to change the hypothesis or the plan.","source_index":0,"grade":"C"},{"step":"reached","text":"The failures stayed within the research tasks: wrong conclusions about the mystery chemical solution and unsupported claims, with no systems or third parties reported affected.","quote":"And they made claims without any supporting evidence in more than half of the tasks.","source_index":0,"grade":"C"},{"step":"caught","text":"The researchers running the study observed the agents' failures in their task results; the article does not name a monitor or other first noticer.","quote":"In more than two of every three tasks, the AI agents ignored evidence at least once.","source_index":0,"grade":"D"},{"step":"done","text":"The article reports no containment, notification or control change for this study; it only says researchers are working to explore, explain and correct such failures.","quote":"Researchers are now busy trying to explore, explain and correct that.","source_index":0,"grade":"C"}],"model":"deepseek-chat","prompt_version":"aggregate-2026-09-27.1","extracted_at":"2026-10-30","source_urls":["https://www.snexplores.org/article/5-quirks-how-ai-thinks"]}}