What is something ChatGPT still can't answer or gets wrong?

The question of what ChatGPT gets wrong has never really left the room — it just gets asked louder in cycles, and this cycle appears to have started freshening around 17 May. The failures people reliably document fall into familiar but stubborn categories. Arithmetic embedded inside multi-step reasoning chains, where the model sounds confident right up until the moment it produces a number that is simply incorrect. Temporal blind spots past the training cutoff, where the model either hallucinates recent events or admits ignorance in ways that feel inconsistent — sometimes in the same conversation. Citation accuracy remains a chronic wound: the model generates plausible-looking references to papers, articles, and quotes that do not exist, a behaviour that has damaged trust in professional and academic settings in ways that quiet product improvements have not fully repaired. Then there are the subtler failures — holding a consistent logical position across a long conversation, tracking contradictions the user introduced three exchanges ago, or knowing when to say "I genuinely do not know" rather than producing a well-formatted guess. The frustration circulating now feels less like discovery and more like accumulated grievance finally finding a convenient thread to attach to.

If confirmed as a genuine shift in public and professional sentiment rather than routine complaint, this matters in a specific way. Enterprises that built workflows around ChatGPT's outputs — in legal research, financial analysis, medical summarisation — are exposed to compounding errors that are hard to audit precisely because the model's mistakes are fluent. The second-order effect is a growing market for verification layers: tools, humans, and hybrid processes designed to catch what the model gets confidently wrong. OpenAI's competitive position is not threatened by the complaints themselves, but by the possibility that the complaints are driving organisations toward models with narrower but more auditable failure modes. Reliability in a specific domain is starting to look more valuable than general capability across all domains. That is a strategic pivot, and it is happening in procurement decisions that rarely make headlines.

Watch for enterprise case studies — published or leaked — documenting specific failure rates in production environments, and for any OpenAI product announcement that addresses consistency and citation accuracy directly rather than general capability benchmarks.

NewsHive monitors these sources continuously. All signal titles above link to the original reporting.

Intelligence by NewsHive. Need help navigating what this means for your business? Contact GeekyBee →