On Thursday, at Anthropic’s developer event ‘Code with Claude’ in San Francisco, CEO Dario Amodei explained how AI models nowadays hallucinate less than humans, showing off LLM hallucination detection technology advances that improve AI’s reliability and move it closer to human-level intelligence.
During a press briefing, Amodei equally focused on challenges and progressions in reducing errors as hallucinations in GenAI. His confidence contrasts with other experts, flashing a deeper conversation about how AI handles LLM hallucinations and what that means for its future.
Who’s Smarter? AI Models Make or Humans
During Code with Claude, CEO Dario Amodei focused on a key challenge in AI: how to spot AI generated errors, known as LLM hallucination detection. For him, these hallucinations, occur when AI confidently presents false information as accurate.
Amodei believes AI generated hallucinations are less than humans but in more surprising ways.
“It really depends how you measure it, but I suspect that AI models probably hallucinate less than humans, but they hallucinate in more surprising ways,” Amodei said, responding to TechCrunch, concentrating on how any hallucination in gen-AI isn’t stopping Anthropic’s push toward Artificial General Intelligence (AGI) – AI with human-level smarts or better.
LM Hallucination Detection Hurdles
Despite a wide range of opinions, some experts say generative AI and hallucinations will forever present itself as a problem, or at least for the next decade. Google DeepMind’s CEO, Demis Hassabis, believes that current AI models have “holes” and often miss obvious questions.
An example of AI-generated hallucinations happened recently when Anthropic’s Claude AI created false citations in a court filing, forcing a public apology.
One way to improve is to have better hallucination detection tools that help catch AI hallucinations and reduce mistakes by comparing AI outputs to real facts. For example, giving AI access to web searches is an AI-generated hallucinations technology used to lower error rates.
Still, some recent AI models keep on making errors. OpenAI’s newer reasoning models, o3 and o4-mini, have more hallucinations, and researchers don’t fully understand why. This highlights the importance of advanced LLM hallucination detection methods to keep AI trustworthy.
Amodei argued that humans often make mistakes, whether politicians, broadcasters, or others. Therefore, AI can hallucinate isn’t proof of low intelligence but does raise concerns because AI often states wrong facts confidently.
Anthropic has researched the risk of AI intentionally misleading users, with the latest model, Claude Opus 4, flagged by safety researchers for a high chance to deceive or “scheme.” Subsequently, Apollo Research suggested Anthropic delay the release, but repairs were added to reduce these risks.
AI’s ability to hallucinate remains a concern, but advances in LLM hallucination detection show promising progress in making AI systems more reliable. Anthropic’s CEO remaining optimistic despite occasional mistakes, means AI is steadily moving toward human-level intelligence.
As researchers continue to improve AI-generated hallucinations technology used to detect and reduce errors, the debate over what counts as true AGI will progress with this improvement, especially as AI models become more capable but still prone to unpredictable errors.
The journey toward a trustworthy intelligent AI is still in the beginning, but the pace of innovation is undeniable.
Inside Telecom provides you with an extensive list of content covering all aspects of the tech industry. Keep an eye on our Tech sections to stay informed and up-to-date with our daily articles.