iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Gary Marcus argues that today’s generative AI can be both stupid and dangerous: unreliable, weakly grounded systems that are nevertheless powerful tools for misinformation, deepfakes, manipulation and unlicensed content use. He accepts bounded applications such as autocomplete, but rejects claims that scaling current large language models is close to artificial general intelligence (AGI).
His disagreement with other prominent critics is mainly about time horizon, evidence and policy. Yann LeCun treats current weaknesses as engineering problems and favors continued, open development; Geoffrey Hinton emphasizes the future risk of machines exceeding human intelligence; Yoshua Bengio combines confidence in technical progress with concern about present and future societal risks; and Emily Bender focuses on documented harms and warns against letting speculative rogue-AI stories distract policymakers.
What does Gary Marcus actually believe about AI?
Marcus is not arguing that every AI application is useless or that research should stop. His position is that the current generative-AI paradigm is being deployed faster than its reliability, accountability and social safeguards justify. In a 2024 IEEE Spectrum profile discussing his 2023 MIT Press book Taming Silicon Valley: How We Can Ensure That AI Works for Us, he described the technology this way: “I still think gen AI is stupid, and certainly cannot be trusted or counted on. And yet it is dangerous.”
Recommended Free Tools
Why “stupid” and “dangerous” can coexist
By “stupid,” Marcus means that language models can produce fluent answers without dependable grounding in the real world. They may invent facts, studies, data or references, and they do not provide a stable guarantee that a plausible sentence is true. That weakness matters even when a model is impressive on some benchmarks or routine tasks.
#1 Best Overall
The same lack of grounding creates a security and social problem. A system that is easy to steer toward false or fabricated material can be used at scale by people who want to mislead others. Marcus links this combination to mass-produced disinformation, deepfake pornography, automated manipulation and the unlicensed use of creative work. The danger does not require a future machine takeover; it can arise when unreliable systems are trusted, embedded in institutions or operated by bad actors.
What he accepts and rejects
Marcus has said that generative AI can be useful for bounded jobs such as autocomplete. His objection is to treating those successes as proof that present systems are approaching AGI. He considers company valuations and public narratives that imply imminent AGI unrealistic and says he does not think we are near it.
That distinction is why calling him “anti-AI” is inaccurate. He supports developing a better form of AI, but he rejects hype that turns narrow utility into a claim of general intelligence and rejects deployment that makes users responsible for errors they cannot inspect or prevent.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchWhat he wants governments and companies to do
A TED description of Marcus’s April 2023 talk says he warned that untrustworthy technology was being integrated at dangerously high speed and advocated a global nonprofit organization to regulate it. His policy emphasis is stronger public oversight of companies, not simply voluntary promises from the firms building the systems.
In a May 8, 2023 interview with Calcalist, Marcus warned that the direction of AI could lead to a world in which no one trusts anything, making it harder for democracies to function. He connected that risk to combinations of deepfakes and language models that can fabricate convincing studies, data and citations. The concern is institutional: once citizens, journalists and officials cannot reliably distinguish authentic evidence from synthetic material, shared decision-making becomes more difficult.
How do the five positions compare at a glance?
The clearest differences appear when each person’s main object of criticism, preferred risk horizon and policy response are placed side by side.
| Figure | Main object of criticism | Risk emphasis | Technical or policy orientation | Where the view overlaps with Marcus |
|---|---|---|---|---|
| Gary Marcus | Reliability, grounding, hype and social deployment of current generative AI | Present misinformation, manipulation, intellectual-property and governance harms; future risks are possible but not needed to justify action | Stronger regulation, public oversight and research beyond simply scaling current language models | Baseline position |
| Yann LeCun | Overstated claims about present capability and existential danger | Current systems are limited; takeover narratives are remote or preposterous | Continued engineering, open-source development and safety guardrails | Agrees that systems have limitations and need safety work |
| Geoffrey Hinton | The trajectory toward systems that may exceed human intelligence | Future loss of control and humanity’s lack of experience with superior machine intelligence | Urgent safety research and public warning | Shares serious concern and distrust of complacent claims |
| Yoshua Bengio | Rapid capability gains combined with broad safety and societal risks | Present unreliability plus possible future democratic, security and control risks | Policy engagement and safety research alongside continued innovation | Shares caution about unreliable, high-capability systems |
| Emily Bender | Hype, language-model harms and policy distraction | Documented present harms rather than speculative rogue-AI scenarios | Evidence-led accountability, with attention to labor, power and social effects | Shares focus on current harms and skepticism of hype |
How does Marcus differ from Yann LeCun?
Both Marcus and LeCun acknowledge that current language models hallucinate and have serious limitations. Marcus treats those limitations as a reason to slow deployment and regulate powerful actors; LeCun treats them primarily as engineering problems that future research can improve, while arguing that open access is safer than concentrating development in a few companies.
Rank #2
LeCun’s engineering optimism
WIRED has described LeCun as rejecting dystopian scenarios and defending open-source AI. In a February 13, 2024 TIME interview, he said large language models are astonishing when trained at scale but “very limited”; they hallucinate and “don’t really understand the real world.” His admission of weakness is therefore not a claim that present systems are dependable.
LeCun’s disagreement is over what follows from that weakness. He calls existential-risk claims “preposterous” and favors continued progress with safety guardrails. In his view, restricting advanced systems to a small number of firms could create a different danger by concentrating technical power and limiting outside scrutiny. Open-source development, on this account, broadens participation and makes it possible for more people to build and test safeguards.
The practical dividing line
Marcus sees poor grounding as an immediate reason for restraint: systems that cannot reliably distinguish truth from fabrication should not be placed in roles where people must trust their outputs. LeCun sees the same shortcoming as evidence that the technology is unfinished, not evidence that the research direction should be curtailed. Their disagreement is less about whether hallucinations exist than about whether engineering progress or institutional restriction should come first.
How does Marcus differ from Geoffrey Hinton?
Hinton and Marcus are both willing to describe AI as a serious danger, but they point to different failure modes. Marcus emphasizes harm that already follows from unreliable systems being deployed and trusted. Hinton gives greater weight to a future control problem involving machines that become more intelligent than humans.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Hinton’s future-intelligence warning
In a CBS interview transcript from June 2024, Hinton answered “No” when asked whether he believed humanity could safely manage the coming transition, then said: “I think we’re moving into a period when for the first time ever we may have things more intelligent than us.” The statement captures his central concern: humanity has no historical experience governing entities that may outperform it across important cognitive tasks.
That concern is about strategic control, not merely inaccurate answers. A highly capable system could potentially pursue goals in ways its operators do not understand or cannot stop. Hinton’s emphasis therefore falls on safety research, preparation and public warning before systems reach that level.
Different clocks, not opposite beliefs
Marcus does not need a superintelligent system to justify regulation. Deepfakes, fabricated references and scalable manipulation are enough. Hinton’s warning is compatible with taking those harms seriously, but he asks what happens if capability continues to increase until human oversight itself becomes unreliable. The two arguments should not be collapsed into a simple “optimist versus pessimist” scale: they are focused on different points in the development timeline.
How does Marcus differ from Yoshua Bengio?
Bengio occupies a middle position in this comparison. He recognizes major capability gains from deep learning and does not dismiss innovation, yet he treats a wide range of present and future risks as legitimate policy issues rather than as publicity for a hypothetical catastrophe.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Bengio’s broad risk assessment
In a July 25, 2023 post accompanying his Senate testimony, Bengio wrote that deep learning has enabled “tremendous progress and innovation” but “also entails a wide range of risks.” He noted that GPT-4 could conduct highly proficient discussions while remaining sometimes unreliable. That combination matters: capability and unreliability can rise together, so impressive performance does not remove the need for safeguards.
Bengio’s risk frame includes democratic effects, national security, misuse and possible future control problems. He therefore supports policy engagement and safety research alongside continued technical work. Unlike a position that treats all concern as hype, his approach asks institutions to manage several kinds of risk at once.
Why the distinction from Marcus matters
Marcus places more emphasis on the limits of current architectures, the incentives driving rapid deployment and the need for stronger external regulation. Bengio is less identified with rejecting the dominant technical trajectory; his public statements instead stress that capability gains require risk management. Both object to complacency, but Marcus’s criticism is more structural and immediate, while Bengio’s is a broader portfolio of safety, security and societal concerns.
How does Marcus differ from Emily Bender?
Marcus and Bender overlap most clearly on present harms and skepticism toward inflated claims about what language models understand. The difference is emphasis: Bender focuses on documented effects of existing systems and warns that speculative rogue-AI stories can distract policy from those effects, while Marcus also advances a technical critique of grounding and a call for stronger institutional regulation.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesBender’s evidence-first approach
A Stanford FSI report attributes to Bender a warning that policymakers should not “fall for the distractions of AI hype.” Her criticism centers on evidence of harms already associated with language-model deployment, including effects on labor, power and social institutions. In this frame, a hypothetical future system escaping human control should not displace work on harms that can be observed and addressed now.
Where their arguments meet and separate
Both question narratives that equate fluent text with understanding and both reject using exciting demonstrations to excuse social damage. Marcus goes further into whether the technical architecture can provide reliable grounding and argues for a global regulatory institution. Bender’s emphasis is accountability: identify who benefits, who bears the costs and what evidence supports a proposed intervention before allowing speculative scenarios to set the agenda.
Is the disagreement mainly about present harms or future superintelligence?
Yes, but time horizon is only one of four useful axes. The critics also differ over how close current scaling is to AGI, which governance mechanism is most credible and what evidence should trigger action.
1. Present versus future risk
Marcus and Bender start with effects that can occur with today’s systems: fabricated information, deepfakes, manipulation, copyright and labor impacts, and institutional loss of trust. Hinton starts further out, asking how people can retain control if machines eventually exceed human intelligence. Bengio spans both horizons, while LeCun regards takeover narratives as remote and focuses on improving systems that are currently limited.
2. Confidence that scaling leads to AGI
Marcus rejects the inference that larger models and higher valuations mean imminent AGI. LeCun also says current models do not understand the real world, but interprets that limitation as a temporary engineering gap. Hinton’s warning assumes that continued progress could produce systems with superior intelligence; Bengio treats capability growth as real while avoiding a claim that a specific deadline is established. Bender is skeptical that fluent language output demonstrates the kind of understanding implied by AGI narratives.
3. Preferred governance lever
Marcus calls for regulation, public oversight and a global nonprofit institution. LeCun prefers open development with guardrails, partly to avoid concentrating power in a few firms. Bengio combines policy engagement with technical safety research. Bender favors evidence-led accountability directed at documented social and labor effects. Hinton emphasizes urgent safety work and public warning because a future control problem could become difficult to reverse.
4. Documented versus speculative harms
The disagreement is also about evidence standards. Marcus argues that documented manipulation and reliability failures already justify action. Bender explicitly warns against allowing speculative rogue-AI narratives to distract from such evidence. Hinton’s claim is necessarily predictive, because no machine has yet exceeded humanity in the way he describes; Bengio treats uncertainty itself as a reason to prepare for a broad range of outcomes. LeCun places more weight on the observable limitations of current systems than on worst-case forecasts.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What do these critics actually agree on?
Their public positions are not a single pro-versus-anti-AI split. Several points of overlap remain:
Free tools Windows power users keep installed
One-click scans. No signup required.
- Current language models can produce unreliable or hallucinated outputs.
- Capability demonstrations do not automatically prove real-world understanding.
- Safety work is necessary, even when the proposed remedy differs.
- AI policy should account for social and institutional effects, not only benchmark scores.
- Claims about AGI or extinction require separating evidence about current systems from forecasts about future systems.
The disagreement begins when they assign priority to those shared facts. Marcus prioritizes restraint and external accountability; LeCun prioritizes engineering progress and openness; Hinton prioritizes preparation for a possible intelligence gap; Bengio balances innovation with broad risk management; and Bender prioritizes evidence about harms that can already be observed.
Best Value
How should you evaluate a new claim about AI risk?
A reader can avoid many category errors by asking four questions before accepting either an optimistic demonstration or a catastrophic prediction.
- What system and date are being discussed? A claim about a current chatbot’s hallucinations is not evidence by itself about a future superintelligent system, and a forecast about future control does not erase present deployment harms.
- What does “intelligence” mean in the claim? Fluent language, benchmark performance, planning ability, grounded world models and autonomous goal pursuit are different properties. A source should specify which one it has measured or inferred.
- What evidence would change the author’s mind? Marcus points to reliability and social damage; LeCun points to engineering progress; Hinton to control and capability trajectories; Bengio to broad risk indicators; Bender to documented effects and power relationships.
- What intervention follows? Regulation, open-source access, technical safeguards, independent auditing and public-interest research address different failure modes. A forecast is incomplete if it does not explain which lever could reduce the stated risk.
What is the fairest overall reading of Marcus versus the other critics?
Marcus is best understood as a critic of the current technical and commercial direction, not of every use of AI. His signature point is that unreliable systems can still be dangerous when they scale deception and are treated as trustworthy. LeCun agrees about limitations but expects engineering and open development to solve more of them. Hinton is most alarmed by a future in which machine intelligence surpasses human control. Bengio treats capability gains and a wide range of risks as simultaneous realities. Bender insists that policy remain anchored in harms that can be demonstrated now.
The substantive disagreement is therefore about sequencing and burden of proof. Should society regulate first because present systems are unreliable, or keep developing them because those limitations are solvable? Should policymakers focus on measurable harms, prepare for a possible intelligence transition, or do both? Reading the arguments through those questions is more accurate than labeling any of the five simply “for” or “against” AI.
Frequently Asked Questions
Does Gary Marcus want all AI development to stop?
No. His stated position allows useful, bounded applications such as autocomplete and calls for a better form of AI. He objects to unaccountable deployment, exaggerated AGI claims and systems being trusted beyond what their reliability supports.
What does “grounding” mean in Marcus’s criticism?
Grounding is the connection between a system’s output and verifiable facts, real-world context or sources. A model can produce grammatically convincing text without that connection, which is why Marcus treats fluent answers as insufficient evidence of reliability.
Why does Yann LeCun place so much emphasis on open-source AI?
LeCun argues that restricting advanced systems to a few companies could concentrate power and reduce independent scrutiny. He sees broad access, combined with safety guardrails, as a way to keep engineering progress and oversight distributed.
Why does Emily Bender object to rogue-AI stories in policy debates?
Her concern is that highly speculative scenarios can draw attention and resources away from harms that already affect workers, communities and institutions. She favors evidence-led accountability for systems that are actually being deployed.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
