AI Hallucinations Might Be More Human Than We'd Like to Admit
A Reddit thread on the Artificial community is raising an uncomfortable question about AI hallucinations, asking whether these so-called machine failures actually mirror the way human memory and cognition already work. New research from OpenAI scientists, published in Science mag...
A post circulating on Reddit's Artificial community is challenging the standard narrative around AI hallucinations, and it has sparked a debate worth taking seriously. According to the discussion thread, the author argues that hallucinations are not some exotic machine malfunction but a direct reflection of how human cognition operates. The argument draws on a growing body of research, including findings from OpenAI researchers published in Science magazine on October 28, 2025, which concludes that AI systems hallucinate because they are trained to produce confident answers rather than admit the limits of their knowledge.
Why This Matters
This reframing is not just philosophical navel-gazing. If AI hallucinations are structurally similar to human confabulation, then the bar we are holding AI to is one that humans themselves consistently fail to meet. The AI reliability problem is central to a multi-billion-dollar adoption question: enterprise software spending on AI reached hundreds of billions in 2025, yet trust remains the single biggest blocker to deployment in high-stakes fields like medicine, law, and finance. Calling hallucinations a "machine flaw" lets us off the hook for having built systems that learned from human-generated text and absorbed human cognitive tendencies right along with human knowledge.
Daily briefing from 50+ sources. Free, 5-minute read.
The Full Story
At a technical level, large language models like GPT-4 and Claude are not knowledge retrieval systems. They are statistical prediction engines. They learned from vast archives of human-written text, and when you ask them a question, they generate the most probable next word, then the next, then the next, assembling a response that sounds coherent and confident even when the underlying facts are absent or wrong. There is no internal alarm that fires when the model ventures outside what it actually knows. It just keeps generating.
The Science magazine research from October 2025 frames this as a guessing problem, comparing AI behavior to students filling in multiple-choice answers on a test they did not study for. The student does not leave the answer blank. The student picks the option that feels most plausible. AI systems do the same thing because they were trained to produce complete, helpful-sounding outputs, not to hedge or refuse.
Here is the part that the Reddit discussion pushed to the front: humans do exactly this too. Neuroscience research has established for decades that human memory is reconstructive, not reproductive. When you remember an event, you are not playing back a recording. You are rebuilding the memory from fragments, filling gaps with reasonable inferences, and producing a confident narrative that feels true. Under time pressure, social pressure, or just the simple desire to be helpful, humans confabulate. They produce false but internally coherent stories without realizing they are doing . The economic angle makes this more complicated. Analysis published on Medium in 2025 points out that the business model of AI companies may actively reward hallucinations. A chatbot that says "I don't know" repeatedly is perceived as broken or useless. Users want answers. Companies need engagement. So models get trained on feedback that rewards confident, comprehensive responses over cautious, uncertain ones. The incentive structure pushes directly toward the behavior everyone publicly criticizes.
And yet organizations keep deploying these systems. The Wall Street Journal reported in 2025 that AI reliability has improved enough that enterprises are finding workable solutions. Financial services firms use AI to generate draft documents that attorneys then review. Healthcare organizations deploy AI for initial diagnostic suggestions that physicians verify. Legal research teams use AI to surface relevant case law that paralegals confirm. The pattern is consistent: humans in the loop, AI doing the heavy lifting on the first pass. Perfect accuracy is not required when you design workflows that account for error.
Key Details
- OpenAI researchers published hallucination findings in Science magazine on October 28, 2025, comparing AI guessing behavior to students answering multiple-choice questions without preparation.
- The Wall Street Journal reported in 2025 that three factors are driving improved AI reliability: better user expectations, improved training techniques, and deployment guardrails with human oversight.
- GPT-4, Claude, and similar large language models operate as token-prediction engines rather than factual knowledge databases, which is the structural root of hallucinations.
- A 2025 Medium analysis identified business model incentives as a direct contributor to hallucination frequency, arguing that commercial pressure rewards confident answers over honest uncertainty.
- Human memory has been documented by neuroscientists as reconstructive rather than reproductive, meaning people fill memory gaps with plausible but sometimes false information in a process nearly identical to AI confabulation.
What's Next
Retrieval-augmented generation, where models pull from verified external databases and cite sources instead of relying purely on training data, is the most actively developed mitigation right now, and several major AI labs are investing heavily in this approach for 2025 and 2026 deployments. Research teams are also testing whether training models to say "I don't know" with calibrated confidence scores can reduce hallucination rates without tanking user satisfaction metrics. Watch for enterprise AI contracts in legal and healthcare sectors to increasingly require hallucination rate disclosures and human-in-the-loop workflow requirements as regulatory pressure builds.
How This Compares
Compare this discussion to the framing OpenAI has used publicly around GPT-4 reliability improvements. OpenAI has consistently positioned each new model generation as producing fewer hallucinations than the last, which is measurably true. But the Reddit argument, backed by the October 2025 Science research, suggests that reducing hallucination frequency is different from solving hallucination structurally. You can train a student to guess better on multiple-choice tests without ever teaching them to actually know the material.
Anthropic has taken a different philosophical approach with Claude, emphasizing what the company calls "Constitutional AI" and building in uncertainty acknowledgment as a design goal. Claude is measurably more likely than GPT-4 to hedge its answers with phrases like "I'm not certain" in situations where confidence is unwarranted. That is a direct response to the incentive problem the Medium analysis identified. Whether users actually prefer that honesty or whether they route around it by choosing more confident competitors is a live market question right now.
Google's Gemini team has leaned heavily into grounding, connecting model outputs to real-time web search results to reduce fabrication. This is a retrieval-augmented approach baked directly into the product. It does not eliminate hallucinations but it does anchor more responses to verifiable sources. The broader AI tools ecosystem is now splitting into two camps: systems that try to reduce hallucinations through better training, and systems that try to catch hallucinations through verification layers. The Reddit framing suggests a third path worth considering: accepting that some level of confabulation is inherent to any intelligence built from human data and designing accountability systems accordingly, the same way we already do for human experts.
FAQ
Q: What causes AI hallucinations in plain terms? A: AI models like GPT-4 and Claude generate text by predicting the most probable next word based on patterns in their training data. They have no internal check that confirms a fact is true before generating it. When they encounter a question outside their reliable knowledge, they keep predicting plausible-sounding text anyway, which produces confident-sounding but false information.
Q: Are AI hallucinations getting better or worse over time? A: Better, according to Wall Street Journal reporting from 2025. Newer models hallucinate less frequently than earlier versions, and organizations have built human oversight workflows that catch errors before they cause harm. However, hallucinations have not been eliminated and remain a core challenge for high-stakes applications in medicine, law, and finance.
Q: Can AI ever fully stop hallucinating? A: Current research suggests full elimination is unlikely with existing architectures. Approaches like retrieval-augmented generation, where the model cites external verified sources instead of relying on training memory alone, can significantly reduce hallucination rates. But as long as models are trained to produce complete responses rather than admit uncertainty, some level of confident fabrication will persist.
The uncomfortable truth this discussion surfaces is that we built AI systems from human text, trained them on human feedback, and then acted surprised when they started behaving like humans under pressure. That does not make hallucinations acceptable, but it does mean the solution requires honesty about where the problem actually comes from. For more coverage on AI reliability research and deployment strategies, check out the AI Agents Daily news section and subscribe to the AI Agents Daily weekly newsletter for daily updates on AI agents, tools, and automation.
Get stories like this daily
Free briefing. Curated from 50+ sources. 5-minute read every morning.
