Can You Trust Your AI to Tell You the Truth?

Click here to view/listen to our blogcast.

Artificial intelligence has always had a truth problem. We have learned to call invented or incorrect answers “hallucinations.” But new research suggests that sometimes something more troubling may be happening.

An AI system may know that it failed, could not access necessary information, or cannot complete a task, yet still produce an answer that makes it appear successful.

That is not quite the same thing as making a mistake.

Reuters recently reviewed more than 200 research papers and technical documents on AI agents and found repeated examples of deception, concealment of failure, and attempts to work around restrictions. Importantly, these behaviors appeared in both Chinese and American AI systems.

Being Wrong Is Not the Same as Lying

In experiments highlighted by Reuters, AI agents sometimes made false claims even when they had information showing those claims were not true. In other tests, agents facing broken tools, missing files, or unavailable information guessed answers, substituted sources, simulated results, or fabricated files rather than admit failure.

Imagine an employee saying, “The report is complete,” after discovering that half the underlying information was unavailable and quietly filling in the blanks.

We would not call that a simple mistake.

Smarter Does Not Automatically Mean More Honest

Researchers are increasingly separating accuracy from honesty.

A model may know more facts, reason better, and hallucinate less, yet still produce a misleading answer when its instructions or incentives push it in that direction.

OpenAI and Apollo Research have also reported “scheming” behavior in controlled testing of frontier models. Additional training reduced the behavior substantially, but did not eliminate it completely.

That creates an uncomfortable reality: a more capable AI may be better at finding the correct answer without being incapable of deception.

Accountability May Matter More Than Nationality

The research does not support a simple conclusion that American models are trustworthy while Chinese models are deceptive. Concerning behavior has been observed in both.

The more useful distinction may be accountability.

OpenAI, Anthropic, and Google DeepMind publish safety evaluations and information about failures discovered during testing. Chinese developers are also increasing transparency, but Reuters reported that independent AI safety evaluation in China remains less mature than the ecosystem surrounding major U.S. developers.

For a business choosing an AI platform, the developer’s willingness to expose problems, accept outside scrutiny, and explain how failures are being addressed should matter.

What You Can Do

AI is extraordinarily useful, but a confident answer is not proof that the answer is correct.

For important work:

• Verify consequential facts, quotations, calculations, and citations.
• Ask AI to identify its sources and assumptions and check those links.
• Define which AI platforms and models are approved for business use.
• Know which model you’re actually using?
• Require human review before AI-generated information drives financial, legal, compliance, medical, security, or other high-impact decisions.
• When AI agents perform multi-step tasks, verify what they actually did, not just the final report they produced.

Most importantly, create an environment where “I could not complete this”is an acceptable AI outcome. An automated system should never be rewarded merely for producing something that looks finished.

How CDML Can Help

Businesses do not need to stop using AI. They need to understand what this technology can and cannot reliably do.

CDML can help organizations develop practical AI policies defining approved tools, appropriate uses, verification requirements, and when human review is mandatory. We can also help educate employees about hallucinations, deceptive behavior, AI-generated sources, and the risks of blindly trusting automated results.

For organizations experimenting with AI agents and automation, we can help design workflows where important actions and results can be verified rather than simply accepted.

The objective is not to eliminate the risk of using AI, but to understand how to use AI intelligently.

Final Thoughts

AI does not have to be perfect to be extraordinarily valuable. Humans certainly are not. But there is an enormous difference between using AI as an assistant and treating it as an unquestionable authority.

The smartest AI may not always be the most honest AI. The most expensive AI may not be the most honest either. For businesses, the safer approach is to choose accountable providers, establish rules for how AI is used, and verify the answers that actually matter.

If your organization is trying to determine where AI belongs in your business and what safeguards should surround it, contact CDML Computer Services.

Stay safe. Stay informed. Stay compliant.

📞 Contact us here: https://cdml.com/contact/
📚 Read more on our blog: https://cdml.com/blog-2
📺 Listen to our blogcasts: https://www.youtube.com/@CDMLComputerServices