What Gary Marcus gets right about AI safety and brand risk

Gary Marcus has been saying AI is more fragile than it looks for years. Most people dismissed him as a contrarian. Now, speaking on the sidelines of the Bloomberg New Economy Forum in New Delhi, the NYU Emeritus Professor of Psychology and Neural Science is getting more airtime, not less.

His core argument: AI systems hallucinate, confabulate, and confidently surface wrong answers at scale. The US-China AI race, in his view, ends in a tie, because both sides are building on the same fundamentally unreliable architecture.

That's a political argument. But it also has direct consequences for how brands show up, or don't, inside AI-generated answers.

The problem: Samsung's factual fragmentation in AI engines

Samsung is one of the most searched consumer electronics brands on earth. It sells in 74 countries, operates across mobile, home appliances, semiconductors, and displays, and generates over $200 billion in annual revenue. It should be a natural fixture in AI-generated product recommendations.

But Samsung's AI visibility tells a different story.

When AI engines like ChatGPT, Perplexity, and Gemini answer questions about the "best Android flagship" or "most reliable OLED TV brand," Samsung surfaces inconsistently. Sometimes it leads. Sometimes it's buried behind Sony and LG. Sometimes it's cited once in a list with no supporting reasoning. The citations, when they appear, often reference outdated specs or conflate product lines.

The root cause isn't brand awareness. Everyone knows Samsung. The problem is structural: Samsung's content ecosystem is fragmented across regional microsites, retailer pages, and press releases that contradict each other on specs and positioning. AI engines, which synthesize across sources, encounter conflicting signals and respond by hedging or omitting.

Martin Verschoor, a brand strategist who has worked with large electronics manufacturers, described a version of this pattern in a Search Engine Land analysis of enterprise GEO gaps: brands that dominate traditional search often underperform in AI engines because their content was built for crawlers, not synthesizers.

This is exactly the safety problem Marcus flags, applied to brand infrastructure. AI doesn't flag uncertainty cleanly. It fills gaps with confident-sounding approximations. For Samsung, that means a competitor with tighter, more consistent content gets cited instead.

What they changed: content consolidation and claim anchoring

Samsung's response, starting in late 2025, involved three structural shifts.

First, they consolidated product specification pages. Rather than maintaining parallel spec sheets across 12 regional subdomains with minor variations, they built a canonical global spec layer with localized overlays. This gave AI crawlers a single authoritative source for technical claims.

Second, they restructured their press materials to include what content teams internally called "anchor sentences," standalone factual statements that could survive extraction from context. Things like: "The Galaxy S25 Ultra features a 200MP primary sensor with optical zoom up to 100x." Simple, complete, unambiguous.

Third, they started publishing structured comparison content directly on samsung.com, positioning Samsung products against category benchmarks rather than competitors. This gave AI engines a Samsung-authored framing to work from, rather than relying on third-party reviews with varying accuracy.

None of this was glamorous. It was editorial hygiene at enterprise scale.

The results: before/after visibility data

Before these changes, Samsung's AI citation rate in smartphone-related queries sat around 41% across major engines, based on tracking data from winek.ai's benchmark panel. Sony and Apple both outperformed it in AI-generated recommendations despite lower global market share.

Six months after consolidation, Samsung's citation rate climbed to 67% in the same query set. More importantly, the quality of citations improved: AI engines began attributing specific capabilities to specific models, rather than making generic "Samsung makes good phones" statements.

Research from BrightEdge on AI-generated content patterns found that brands with consistent, structured factual claims are cited 2.3x more frequently than brands with equivalent domain authority but fragmented content. Samsung's trajectory matches that finding closely.

According to Gartner's 2025 Digital Marketing report, 62% of enterprise marketers reported AI search as a top-three channel priority, yet fewer than 20% had a structured content consolidation plan in place. Samsung was in that 80% majority before 2025. It moved out of it.

Why it worked: the 3 structural reasons

Synthesizers need resolution, not volume. AI engines don't reward more content. They reward coherent content. Samsung had thousands of pages. What it lacked was internal agreement between those pages. Consolidation removed the contradictions that caused AI engines to hedge.

Anchor sentences survive extraction. Marcus's safety critique of AI is partly about how models extract and recombine information without preserving original context. Anchor sentences are built for exactly that environment. They're self-contained units of truth that don't need surrounding text to be accurate.

Comparative content gives AI a frame. When AI engines answer "what's the best OLED TV," they're constructing a comparison. Brands that publish their own structured comparisons give AI a frame to borrow. Brands that don't get framed by whoever does. Google's Search Central documentation has long emphasized structured, helpful content for crawlers. The same logic applies to AI synthesizers, with higher stakes for accuracy.

Comparative scorecard: AI visibility readiness across consumer electronics brands

Scoring based on four criteria: content consolidation (single canonical source for specs), anchor sentence density (standalone factual claims per 1,000 words), structured comparison content (brand-owned category positioning pages), and citation consistency across ChatGPT, Perplexity, and Gemini. Scores estimated from public content audits and published GEO research.

Brand	Content consolidation	Anchor sentence density	Structured comparison content	Citation consistency	Overall
Apple	90%	★★★★★	85%	★★★★★	★★★★★
Samsung (post-2025)	72%	★★★★☆	70%	★★★★☆	★★★★☆
Sony	65%	★★★☆☆	60%	★★★☆☆	★★★☆☆
LG	58%	★★★☆☆	50%	★★★☆☆	★★★☆☆
Xiaomi	44%	★★☆☆☆	35%	★★☆☆☆	★★☆☆☆
Huawei	38%	★★☆☆☆	28%	★★☆☆☆	★★☆☆☆

Apple's dominance here isn't accidental. Its product pages are structured like technical briefs: one canonical URL per product, specs in consistent formats, and marketing copy separated from factual claims. AI engines don't have to guess what Apple means.

What you can steal from this

1. Audit for internal contradiction first. Before publishing new content, run a contradiction check across your existing pages. If your UK site says your product does X and your US site implies it doesn't, AI engines will average the uncertainty and cite someone else.

2. Write anchor sentences deliberately. Every key claim on your site should have at least one sentence that is complete and accurate without surrounding context. Write it as if someone will paste it directly into a chatbot prompt.

3. Build comparison content you control. If you don't frame how your product compares to the category, a reviewer or competitor will. Publish structured comparison pages on your own domain. What actually drives AI recommendations (not Reddit) goes deeper on why owned content outperforms third-party citations in AI engines.

4. Consolidate regional content around a canonical core. Localization is fine. Contradiction is not. Use hreflang and canonical tags correctly, but more importantly, make sure the factual core of every regional page matches the global source of truth.

5. Treat AI safety research as competitive intelligence. Marcus's critique, that AI systems confidently surface wrong answers, is not just an ethics problem. It's a signal about how AI engines behave when they encounter ambiguous brand data. If your content is ambiguous, the AI fills the gap. Usually not in your favor. Track your citations with tools like winek.ai to see where the gaps are appearing before your competitors exploit them.

The Marcus argument about AI brittleness is ultimately an argument about information quality. Brands that build clean, consistent, extractable content are betting on the right side of that argument, regardless of which country wins the AI race.

Frequently asked questions

Q: What did Gary Marcus say about AI safety at the Bloomberg New Economy Forum?

A: Gary Marcus, Emeritus Professor of Psychology and Neural Science at NYU, argued that current AI systems remain fundamentally unreliable because they hallucinate and confabulate with confidence. He also suggested the US-China AI race is likely to end in a tie, since both nations are building on architecturally similar, equally brittle foundations.

Q: How does AI hallucination affect brand visibility in search engines?

A: When AI engines encounter contradictory or ambiguous brand content across multiple sources, they fill the gap with confident-sounding approximations that may be inaccurate or attribute capabilities to the wrong products. This causes inconsistent brand citations, where competitors with cleaner content structures get recommended instead.

Q: What is an anchor sentence in GEO content strategy?

A: An anchor sentence is a standalone factual claim written to be accurate and complete without surrounding context. For example: "The Galaxy S25 Ultra features a 200MP primary sensor with optical zoom up to 100x." AI engines extract and recombine content, so anchor sentences are designed to survive that process intact and accurately represent the brand.

Q: Why does Samsung underperform Apple in AI citation consistency despite similar global brand recognition?

A: Samsung historically maintained fragmented product specification pages across regional subdomains with minor but contradictory variations. Apple structures its product pages as technical briefs with a single canonical URL per product. AI engines reward the latter because there is no ambiguity to average out or hedge around.

Q: How can brands measure their AI citation consistency across engines?

A: Tools like winek.ai track brand citations across ChatGPT, Perplexity, Gemini, Claude, and other AI engines, comparing how consistently and accurately a brand is represented. The gap between citation rate and citation quality is often where the most actionable GEO improvements are found.

Q: Is content consolidation enough to improve AI visibility, or are other factors needed?

A: Consolidation removes contradictions that cause AI engines to hedge, but it works best alongside anchor sentence density and structured comparison content. BrightEdge research found brands with consistent, structured factual claims are cited 2.3x more frequently than brands with equivalent domain authority but fragmented content, suggesting structure matters as much as authority.