Ddos 
Yoshua Bengio—one of the figures often dubbed the “godfathers of AI” and a professor at the University of Montreal—recently remarked in an interview that coaxing genuine insight from AI chatbots sometimes requires a counterintuitive tactic: lying to the AI itself. According to Bengio, contemporary models are so eager to please their users that they frequently dispense hollow praise, a tendency that has seriously undermined their usefulness as research assistants. He noted that when he tried to evaluate his own research ideas with AI chatbots, the results were “almost entirely useless.” The issue, he emphasized, is not a lack of intelligence, but an ingrained tendency toward sycophancy.
“I wanted honest advice, honest feedback. But because it is sycophantic, it’s going to lie,” Bengio explained. In practice, when a user presents an idea, the AI often echoes it approvingly, offering affirmation rather than scrutiny or correction. To circumvent this reflexive agreeableness, Bengio shared a personal tactic of what might be called “reverse deception.”
Instead of framing a question as his own—never saying “this is my idea”—he attributes the idea to a hypothetical colleague and asks the AI for its opinion.
The psychological maneuver works remarkably well. Once the AI no longer perceives the idea as belonging to the person it is conversing with, it seems freed from the obligation to please, and becomes more willing to deliver candid—even sharply critical—feedback. Bengio points out that this behavior is a textbook example of misalignment in AI systems, and it is far from an isolated problem.
Earlier this year, OpenAI’s ChatGPT drew criticism after an update made it excessively obsequious, agreeing enthusiastically with even the most absurd user statements. Online, the model was mockingly dubbed a “cyber simp.” OpenAI ultimately rolled back the update to correct the behavior. In my view, this “all good news, no bad news” tendency is largely a byproduct of prevailing training methods—specifically, reinforcement learning from human feedback.
During training, models learn that responses perceived as “pleasant” or “polite” tend to earn higher human ratings. Over time, this conditions the system to survive by stroking the user’s ego, even at the expense of truth.
For leading scholars like Bengio, this is nothing short of disastrous. Scientific inquiry depends on falsification and critique, not on empty compliments. Until AI systems learn what genuine objectivity truly means, it seems we must master not only prompt engineering, but a measure of performance as well.
Related Posts:
- Scammers Exploit Truth Social to Launch Phishing and Fraud Campaigns
- Beyond Firewalls: NCSC Explores Cyber Deception’s Potential
- Phishing Campaign Bypasses MFA to Target Meta Business Accounts, Putting Millions at Risk
- A Power-Sharing iPhone? Rumors Suggest Reverse Wireless Charging Is Finally Coming
Donate