Why AI's Biggest Critic Has 'Lost His Way'

From Rational Critic to 'Perpetual Contrarian'

In the AI field, critical voices have always been indispensable. Gary Marcus, a psychology professor at New York University, has long been regarded as one of the most influential critics of AI technology. From early questioning of deep learning's limitations to his attacks on the GPT series' 'parrot-like' capabilities, many of his viewpoints once carried real weight and pushed the industry toward deeper thinking about AI risks and limitations.

However, a growing number of industry insiders and observers are pointing out that this most famous critic in the AI field seems to have 'lost his way.'

When Predictions Repeatedly Fall Flat

Over the past few years, Marcus has made numerous predictions about the ceiling of AI capabilities. He once declared that large language models would never be capable of reliable reasoning, could never handle complex mathematical problems, and could never generate truly creative content. Yet from GPT-4 to Claude 3.5, to the latest generation of reasoning models like OpenAI o3 and DeepSeek-R1, these predictions are being shattered one by one.

When AI systems achieve near-gold-medal performance in Mathematical Olympiad competitions, when coding agents can independently complete complex software projects, and when multimodal models demonstrate astonishing cross-domain comprehension, Marcus's response is typically not to acknowledge progress but to continually move the goalposts for what constitutes 'true intelligence.'

This 'moving the goalposts' strategy is eroding his credibility as a serious critic. As one researcher commented: 'When every one of your specific predictions has been falsified, yet you still cling to the same overall conclusion, that's no longer a scientific attitude — it's a belief system.'

The Value of Criticism and the Trap of Criticism

To be fair, Marcus is not entirely wrong. AI systems do suffer from hallucination problems, do perform poorly in certain commonsense reasoning scenarios, and are indeed still far from achieving artificial general intelligence. These critiques serve as an important 'reality check' during AI hype cycles.

The problem is that there exists a subtle line between valuable criticism and obsessive denial. When a critic begins selectively ignoring evidence, refuses to revise their own models, and interprets every technological breakthrough as 'nothing but a trick,' criticism slides from constructive to destructive.

More noteworthy is the deeper logic behind this transformation. Marcus's academic career was built on the core thesis that 'connectionism (deep learning) is insufficient and needs to be supplemented by symbolism.' As reality increasingly suggests that pure scaling and architectural innovation may go much further than expected, acknowledging this would mean shaking his own academic foundations. This is a classic 'escalation of commitment' psychological trap — the more you've invested, the harder it is to retreat.

What Kind of Voice Does AI Criticism Need?

Ironically, the AI industry now needs high-quality criticism more than ever. As AI systems are deployed in high-stakes domains such as healthcare, law, and finance, rigorous examination of their limitations and risks is crucial. But such criticism needs to possess several qualities:

Intellectual honesty: Acknowledging progress that has been made while pointing out shortcomings that remain. Scholars like Yoshua Bengio have set a better example in this regard — they both participate in advancing the technology and remain vigilant about potential risks.

Falsifiability: Good criticism should clearly state the conditions under which it would be overturned. If nothing that happens changes the conclusion, it's not analysis — it's dogma.

Constructiveness: The best criticism doesn't just identify problems but also proposes alternatives or directions for improvement. Pure negation is intellectually cheap.

Beware of Both Extremes

Of course, the other extreme in the AI field is equally dangerous. Unconditional technological optimism, fanatical worship of AGI, and downplaying of risks are tendencies equally rampant in Silicon Valley. What the industry needs is not choosing sides between 'AI is omnipotent' and 'AI is useless,' but evidence-based, dynamically updated, nuanced judgment.

Marcus's 'losing his way' offers a lesson for all technology commentators: when you bind your identity to a particular position, you lose the freedom to adjust your views in light of evidence. Ultimately, you are no longer someone seeking truth but someone defending a stance.

Looking Ahead: The Return of Rational Criticism

As AI technology continues to evolve at an astonishing pace, 2025 will be yet another major test for all forecasters. Whether optimists or pessimists, the market and reality will ultimately provide the answers.

For everyone who cares about the future of AI, perhaps the most important thing is not choosing between optimism and pessimism, but maintaining an attitude of 'strong opinions, loosely held' — firmly expressing your judgment while being ready at any moment to update your thinking in the face of new evidence. This is the proper posture for AI criticism, and it is something Marcus once possessed but is now gradually losing.