Sam Altman says future AI will evolve in "uncomfortable" ways

If two people with opposite political views ask the same AI the exact same question, whose side should the machine take?

Speaking at Davos, OpenAI CEO Sam Altman offered a blunt answer: future AI models will need to allow extensive individual customization, and "that's going to make a lot of people uncomfortable."

In short, OpenAI is moving away from building a single, universal moral arbiter. Instead, ChatGPT is heading toward a world where different users get completely different answers based on their personal values.

Custom values and uncomfortable truths

Altman made sure to draw an absolute line. If a government or user asks the AI to condone killing all gay people, that is "well out of bounds" and an automatic refusal.

Beyond extreme red lines, however, lies the murky territory of cultural and personal beliefs. Altman admitted that while there are views he personally disagrees with, tool builders must accept being uncomfortable with some of the ways their technology gets used.

To Altman, differences between individual users' values matter far more than geographic borders. His vision is a system that adapts to how you think, not just how Silicon Valley thinks.

That aligns with his recent conversation on Bill Gates' podcast, where Altman predicted that reasoning, reliability, and deep personalization powered by personal data will be the defining milestones of the next two years.

GPT-5, cheaper tokens, and robotics

Altman also laid out what OpenAI is building behind the scenes. His top priority right now remains shipping the next flagship model, likely called GPT-5.

Under the hood, the cost of intelligence is collapsing faster than Moore's law. According to Altman, OpenAI has already slashed GPT-3 costs by a factor of 40 and GPT-3.5 by a factor of 10, with the ultimate goal of driving intelligence costs close to zero.

Despite operating with a lean staff of around 500 people, the company's ambitions are stretching out. Altman hinted at returning to robotics by pairing future vision-language models with physical machines.

He also defended recent controversial moves, including opening up OpenAI's technology to military use for the U.S. government, while brushing off the New York Times copyright lawsuit by claiming OpenAI never needed NYT data to build top-tier models anyway.

Why it matters

Altman's embrace of "uncomfortable" customization looks like a calculated move to dump OpenAI's biggest political liability directly into the user's lap.

For the past year, OpenAI has taken flak from both sides of the culture war, accused of being either hopelessly biased or aggressively woke. By letting users tweak the model's value system, Altman neatly sidesteps the impossible task of defining universal truth.

For everyday users and enterprises, this will make AI significantly more practical. You will no longer have to spend dozens of tokens wrestling a model out of its corporate PR tone to get the perspective you actually need.

However, the downside is massive. Perfect personalization is also the perfect recipe for hyper-tailored echo chambers. Outsourcing morality to the user helps OpenAI dodge censorship debates, but when customized tools inevitably end up validating fringe ideologies, simply claiming that tool builders must tolerate discomfort will not shield them from regulatory fallout.