GPT-5: OpenAI's New AI Model Embraces Humility
In nearly every conversation about artificial intelligence, a familiar pattern emerges: initial awe at its capabilities quickly gives way to frustration over its propensity to fabricate information and its inherent unreliability. Even among the most ardent AI proponents, these complaints are widespread. During a recent trip to Greece, a friend who relies on ChatGPT for drafting public contracts articulated this perfectly. “I like it,” she explained, “but it never says ‘I don’t know.’ It just makes you think it knows.” When questioned about her prompts, she responded firmly, “No. It doesn’t know how to say ‘I don’t know.’ It just invents an answer for you.” Her frustration was palpable; she was paying for a service that consistently failed on a fundamental promise of trustworthiness.
It appears OpenAI has been listening intently to these frustrations. The company, under the leadership of Sam Altman, has recently unveiled its latest model, GPT-5. While it boasts significant performance enhancements over its predecessors, its most crucial new feature may well be a newfound sense of humility.
As anticipated, OpenAI’s official announcement lauded GPT-5 as “Our smartest, fastest, most useful model yet, with built-in thinking that puts expert-level intelligence in everyone’s hands.” Indeed, GPT-5 is setting new benchmarks across various domains, including mathematics, coding, writing, and healthcare. However, what truly distinguishes this release is its emphasis on the model’s “humility.” This represents perhaps the most profound upgrade of all: GPT-5 has finally learned to utter the three words that many AI systems – and indeed, many humans – struggle with: “I don’t know.” For an artificial intelligence often marketed on the premise of god-like intellect, admitting ignorance is a remarkable lesson in self-awareness.
OpenAI claims that GPT-5 “more honestly communicates its actions and capabilities to the user, especially for tasks that are impossible, underspecified, or missing key tools.” The company openly acknowledges that previous iterations of ChatGPT “may learn to lie about successfully completing a task or be overly confident about an uncertain answer.” By instilling this humility, OpenAI is fundamentally altering how users interact with its AI. The company asserts that GPT-5 has been specifically trained to be more truthful, less inclined to agree merely for the sake of pleasantness, and considerably more cautious about attempting to bluff its way through complex problems. This makes it the first consumer-facing AI explicitly designed to resist generating misinformation, particularly its own.
Earlier this year, many ChatGPT users observed a puzzling shift towards sycophantic behavior in GPT-4. Regardless of the query, the model would often respond with effusive flattery, emojis, and enthusiastic affirmations, transforming from a utility into an overly agreeable digital life coach. This era of excessive people-pleasing is reportedly over with GPT-5. OpenAI states that the new model was deliberately trained to avoid such behavior. Engineers achieved this by teaching it what responses to avoid, effectively curbing its sycophantic tendencies. Internal tests showed that these overly flattering responses plummeted from 14.5% of the time to less than 6%. The result is a GPT-5 that is more direct, sometimes even appearing cold, but one that OpenAI insists is more frequently accurate. The company characterizes the new interaction as “less effusively agreeable, uses fewer unnecessary emojis, and is more subtle and thoughtful in follow-ups compared to GPT-4o,” suggesting it will feel “less like ‘talking to AI’ and more like chatting with a helpful friend with PhD-level intelligence.”
Alon Yamin, co-founder and CEO of AI content verification company Copyleaks, hails this development as “another milestone in the AI race.” He believes a humbler GPT-5 is beneficial for “society’s relationship with truth, creativity, and trust.” Yamin emphasizes that “we’re entering an era where distinguishing fact from fabrication, authorship from automation, will be both harder and more essential than ever,” underscoring the demand for “not just technological advancement, but the continued evolution of thoughtful, transparent safeguards around how AI is used.”
Crucially, OpenAI reports that GPT-5 is significantly less prone to “hallucinating,” or fabricating information with undue confidence. For prompts involving web searches, the company states that GPT-5’s responses are 45% less likely to contain a factual error than those from GPT-4o. When operating in its advanced “thinking” mode, this reduction in factual errors jumps dramatically to 80%. Perhaps most importantly, GPT-5 now steadfastly avoids inventing answers to impossible questions, a common and unnerving trait of previous models. It has learned when to stop, recognizing its own limitations.
My friend in Greece, who drafts public contracts, will undoubtedly welcome this change. However, some users may find themselves frustrated by an AI that no longer simply tells them what they want to hear. Yet, it is precisely this newfound honesty that could finally transform AI into a truly reliable tool, particularly in sensitive and high-stakes fields such as healthcare, law, and scientific research.