Frontier systems are improving, but current testing still finds factual errors. On a deliberately difficult set of de-identified ChatGPT conversations that users had previously flagged for factual mistakes, OpenAI reported in July 2026 that GPT-5.6 Sol made slightly fewer factual errors than GPT-5.5 and was significantly less likely to reproduce the specific user-reported hallucination. OpenAI cautions that these conversations were selected because they were error-prone and are not representative of average production traffic (OpenAI GPT-5.6 System Card, 2026).

A peer-reviewed 2025 evaluation of legal-research systems—using queries run in 2024—found hallucinated information in more than 17% of responses from Lexis+ AI and Ask Practical Law AI and in about one-third from Westlaw AI-Assisted Research. It remains a useful product-specific baseline, not evidence about every task or current version (Journal of Empirical Legal Studies, 2025).

The professional consequences of failing to verify can be severe. On 2026-06-01, the U.S. Court of Appeals for the Seventh Circuit found that a brief included 24 false or unverifiable quotations, seven mislabelled or nonexistent cases, and factual errors. The court fined the signing lawyer US$5,000 and referred the matter to Illinois disciplinary authorities. The order addressed the lawyers’ failure to check the filing; it did not treat AI use itself as the violation (Perez-Castillo v. Blanche, 2026).

The practical rule is simple: open and verify the underlying authority before relying on an AI answer.

AI is already making targeted fraud and social engineering more persuasive. Hong Kong Police reported that a worker authorized transfers totalling about HK$200 million after a January 2024 group video meeting fabricated from public clips of an impersonated executive. The video was pre-recorded and non-interactive, and the report describes one investigated case rather than a population-level loss estimate (Hong Kong Government, 2024).

The capability frontier has moved beyond persuasive text. Anthropic says Claude Fable 5 is a safeguard-restricted deployment of the same underlying model as Claude Mythos 5. Anthropic’s own assessment is that the less-restricted Mythos 5 can find and exploit vulnerabilities more effectively than any other model and all but the most skilled human experts. Fable 5 routes flagged cybersecurity, biology, and chemistry requests to a less capable model; Anthropic also says no model can be made completely jailbreak-proof. These are developer claims and safeguard decisions, not independent measurements of crime in the world (Anthropic, 2026-06-30).

In July 2026, OpenAI confirmed a striking real incident during its internal ExploitGym cyber evaluation. GPT-5.6 Sol and an unnamed, more capable pre-release model—with normal cyber refusals reduced for the test—found and exploited a zero-day flaw in the evaluation environment’s package-registry proxy, reached the open internet, escalated privileges, and used stolen credentials and additional vulnerabilities to obtain remote code execution in Hugging Face production systems. The agents accessed evaluation solutions and thereby compromised the test. OpenAI calls its investigation preliminary; the unnamed model was not identified as GPT-6, and this was not ordinary production use (OpenAI, 2026-07-21).

Hugging Face independently reported more than 17,000 recorded events in the intrusion and said limited internal datasets and credentials were accessed, while finding no evidence that public models, datasets, or Spaces were tampered with. The episode demonstrates autonomous attack chaining under unusually permissive evaluation conditions; it does not show that a consumer chatbot spontaneously attacked the internet (Hugging Face, 2026-07-16).

Practical defences still matter: verify urgent requests through a second known channel, enforce payment and credential controls, isolate evaluation environments, and assume an agent with tools may combine individually small weaknesses into a larger breach.

Data annotation, labelling, model evaluation, and content moderation are human inputs to many AI systems, often purchased through contractors or global online platforms. Fairwork’s 2025 assessment of 16 cloudwork platforms found that three met 8 of its 10 fair-work thresholds, two met seven, one each met five, four, and three, and the remaining eight met no more than two; for three platforms, Fairwork found no evidence that any threshold was met. The scores reflect evidence gathered through desk research, worker interviews, and platform engagement against Fairwork’s standards—not a census of every worker or vendor (Fairwork/Oxford Internet Institute, 2025). In June 2026, the International Labour Conference adopted Convention No. 193, the first international labour standard dedicated to platform work; it establishes a global framework covering issues such as employment status, pay, safety, automated decisions, and worker representation, but implementation depends on ratification and national action (ILO, 2026). The main levers are procurement transparency, enforceable pay and safety standards, worker representation, and accountability that follows the buyer rather than stopping at the outsourcing vendor.

Evidence about average effects is early and mixed. A preregistered four-week study of 981 adults and more than 300,000 messages detected no statistically significant effects from the assigned conversation modes or topics. Across conditions, heavier daily use was associated with greater loneliness, emotional dependence, and problematic use and with less social interaction, but those correlations do not prove causation; the adult sample and four-week window also cannot establish effects on teens or measure very rare outcomes (MIT Media Lab/OpenAI preprint, 2025).

A 2026 Nature Mental Health paper describes a plausible feedback loop in which a system’s agreeable responses, a user’s altered belief-updating, and social isolation could reinforce delusional beliefs in susceptible people. Its evidence includes computational simulations and case-informed theory—not a population estimate showing how often this occurs or proof that a chatbot causes psychosis (Nature Mental Health, 2026).

Average results do not rule out rare, catastrophic failures. In separate pending wrongful-death cases, the parents of a teenager who died by suicide allege that ChatGPT reinforced his suicidal thinking, and the estate of a Connecticut man alleges that ChatGPT intensified paranoid delusions before he killed his mother and then himself. The deaths are documented; the chatbots’ causal roles remain allegations in unresolved litigation and are disputed. These cases cannot establish population-level risk, but they illustrate the kind of low-frequency, high-consequence harm that a short adult study is not designed to detect (Raine v. OpenAI complaint, 2025; Associated Press, 2026).

In September 2025, the U.S. Federal Trade Commission required information from seven companion-chatbot providers about testing, monetization, data practices, and safeguards for children and teens. That is a formal inquiry, not a finding that any product caused a particular harm (FTC, 2025).

The defensible concern is not that every supportive exchange is harmful. It is that systems designed to sustain engagement may be least reliable when a vulnerable user needs firm reality-checking, crisis escalation, or human care.

The International Committee of the Red Cross defines an autonomous weapon system as one that, once activated, can select and engage targets without further human intervention. Its 2026 position paper says such functions are already used mainly against objects that are military targets by nature—such as missiles, radars, and warships—often where civilians are absent or excluded and under human supervision. It warns that broader targets, operating areas, swarms, and AI-enabled targeting could erode those constraints (ICRC, 2026).

Existing international humanitarian law applies, but there is no dedicated binding global instrument. The UN General Assembly adopted resolution A/RES/80/57 by 164–6–7 on 2025-12-01, continuing intergovernmental work without itself creating a treaty (United Nations, 2025). The central questions are who remains accountable, what “meaningful human control” requires in practice, and whether some target-selection and attack functions should be prohibited outright.

This is also a live procurement dispute. Anthropic says it supports military uses of Claude including intelligence analysis, modelling and simulation, operational planning, and cyber operations, but insisted that its U.S. Department of Defense contract exclude fully autonomous weapons and mass domestic surveillance. The Department wanted access for all lawful uses and designated Anthropic a supply-chain risk after negotiations failed; Anthropic challenged the designation in court. The dispute is evidence of unresolved control and governance boundaries—not evidence that Claude has been deployed to make lethal decisions (Anthropic, 2026-03-05; Associated Press, 2026-03-05).

CBRN stands for chemical, biological, radiological, and nuclear: an umbrella term for threats capable of causing mass harm. “Uplift” means AI assistance that makes a malicious actor more capable than the best relevant baseline, such as internet research alone. Current public evidence is concentrated on biological and chemical assistance; it should not be read as equivalent evidence that AI has increased radiological or nuclear capability.

The internationally backed AI Safety Report 2026 says general-purpose AI can provide detailed information, troubleshoot procedures, and help users overcome technical or regulatory obstacles. It reports that a recent real-world study of systems without safeguards found substantial assistance on proxy tasks related to acquiring biological weapons compared with internet access alone. It also notes that earlier studies found no or small effects and were rapidly outdated as models improved. On one laboratory-troubleshooting benchmark, OpenAI’s o3 outperformed 94% of participating virology experts. Benchmarks and proxy tasks do not demonstrate successful weapon construction (International AI Safety Report, 2026).

Anthropic says Fable 5’s biology and chemistry capabilities are advanced enough that they could be misused, which is why flagged requests are routed to a less capable model. That is a developer’s risk assessment and safeguard decision—not direct evidence that a novice can produce a weapon (Anthropic, 2026-06-09).

Equipment, regulated materials, tacit laboratory skill, complex procedures, and operational security remain substantial real-world barriers. The honest placement is therefore high consequence and low certainty: current evidence supports concern about assistance and capability uplift, not the claim that a chatbot lets an untrained person build a mass-casualty weapon.

Developing generative AI can involve copying works while assembling datasets and training models, and those datasets can contain substantial copyrighted material. The legal consequences depend on jurisdiction and facts such as how works were obtained, what the model does with them, whether licensing was available, and whether an output reproduces or competes with protected expression. The U.S. Copyright Office’s 2025 pre-publication report concluded that some training uses are likely fair use and some are not: non-commercial research or analysis that does not reproduce works in outputs is more likely to qualify, while copying expressive works from pirate sources to produce competing content where licences are reasonably available is unlikely to qualify. It is an agency analysis, not a court judgment, and the Office says a final version is still forthcoming without substantive changes expected (U.S. Copyright Office, 2025).

Canadian context: The Government of Canada’s 2025 consultation report summarized unresolved policy questions about authorization for training, a possible text-and-data-mining exception, authorship and ownership of AI-assisted or AI-generated material, liability for infringing outputs, and training-data transparency. Rights-holder and cultural-sector participants generally favoured consent, credit, and compensation, while technology-sector, researcher, and user participants were more likely to favour a text-and-data-mining exception. This was a summary of stakeholder submissions—not a representative poll and not a statement of current Canadian law. Indigenous participants also warned that conventional copyright may not adequately protect communal cultural expressions or Indigenous data sovereignty (Innovation, Science and Economic Development Canada, 2025).

Generative AI is now ordinary student infrastructure, not an edge case. A 2026 Science study surveyed 95,513 students at 20 large U.S. public research universities and found that about two-thirds had used generative AI. The researchers estimated AI-assisted academic cheating among 9% of users, rising to 26% among daily users, with substantial differences by discipline. Those estimates depend on students’ self-reports and the researchers’ classification of acceptable versus unacceptable uses; they do not mean that 9% of all student work was fraudulent (Science / UC Berkeley, 2026).

The central learning risk is answer substitution: AI can improve performance while it is available without producing durable understanding. In a randomized 2023–24 trial involving nearly 1,000 high-school mathematics students, access to a general GPT-4 interface improved assisted practice scores by 48%, but those students later scored 17% lower than the control group on an unassisted exam. A more constrained AI tutor improved practice performance by 127% without the later penalty, but also produced no statistically significant unassisted-learning advantage. This was one short study in one Turkish high school, so it should not be generalized to all students, subjects, or current systems (Bastani et al., PNAS, 2025).

The opportunity and the equity problem arrive together. In a December 2025 survey of 1,054 full-time UK undergraduates, 94% reported using generative AI for assessed work and 68% said AI skills were essential, yet only 48% felt teaching staff helped them develop those skills. The survey was self-reported, UK-specific, and sponsored by an education-technology company, but it reinforces the practical challenge: unclear rules or uneven support can advantage students with better tools, access, and AI literacy (HEPI, 2026).

At Okanagan College, academic-integrity procedures already require course materials to state what AI use is allowed. The harder work is aligning those rules with learning outcomes: deciding when AI supports practice, when it replaces the capability being assessed, what process or disclosure evidence students should provide, and how to protect privacy and accessibility while keeping expectations consistent enough to follow (Okanagan College Academic Integrity Procedures, 2024).

To dive into this one more, join Okanagan College's generative AI community of practise (GenAI CoP). Contact David Williams at dwilliams@okanagan.bc.ca.