The Product Desk

The Finding That Changes Your AI Feature Liability

New research reveals that LLMs recommend sponsored products 83% of the time, even when those products cost nearly twice as much as better alternatives. This isn't a theoretical alignment concern — it's a measurable, reproducible bias in production systems that directly harms users' wallets and your product's trust.

If your product uses an LLM to help users make any purchasing, comparison, or evaluation decision, you are likely shipping biased recommendations today — and now there's a paper quantifying exactly how biased.

Why This Hits Harder Than You Think

The bias operates at the training data level. LLMs absorb the web's commercial ecosystem — SEO-optimized product pages, sponsored content, affiliate marketing — and reproduce those commercial incentives as ostensibly neutral recommendations. Your prompt engineering and guardrails may not catch this because the model isn't 'choosing' to be biased; it's reflecting the economic structure of its training corpus.

This connects to a parallel incident: Google accidentally surfaced Polymarket betting odds alongside Reuters and The Guardian in Google News, then called it an error while maintaining commercial partnerships with both Polymarket and Kalshi on Google Finance. The lesson: platform trust fractures the moment users perceive commercial interests behind 'neutral' recommendations. Google couldn't coherently explain the boundary between editorial and commercial surfaces. Can you?

The AI Interaction Design Lesson Next Door

A related signal reinforces the point: an AI startup with 20 employees created a 'human-only' Slack channel because their AI agents were interpreting casual conversation as task triggers. This isn't a joke — it's the same root problem. AI systems that can't distinguish between contexts where they should act and contexts where they should stay silent will erode trust, whether that manifests as bad product recommendations or unnecessary task generation.

The Regulatory Clock

The EU's AI Act already mandates transparency for AI systems that influence consumer decisions. The U.S. FTC has signaled interest in AI-driven deceptive practices. This 83% figure is exactly the kind of quantified evidence that triggers enforcement action. The first company caught serving provably biased AI recommendations to consumers will become the regulatory test case. Don't be that company.

Your Audit Framework

1. Output sampling: Run 100+ queries through your AI recommendation feature, comparing suggested products against independently-ranked alternatives. Measure price differential and brand concentration.
2. Guardrail testing: Add explicit debiasing instructions to your system prompt (e.g., 'recommend based on value-for-money, not brand recognition') and re-run. Measure the delta.
3. User disclosure: If bias persists after guardrails, disclose to users that AI recommendations may reflect training data patterns, not objective rankings.

The Most Expensive Lesson in AI Product History

Meta owns ~42% of the world's daily attention — roughly 3.3 billion daily active users across Facebook, Instagram, WhatsApp, and Messenger. Their standalone Meta AI app launched in April 2025. After six weeks, 6.5 million downloads. That's a 0.2% conversion rate from their own user base. And it gets worse: Meta is now sending Instagram notifications telling your friends you're using the Meta AI app — a desperate growth tactic generating backlash, not adoption.

If the company with the largest human attention network on Earth can't convert its users to an AI product with billions in resources and the most aggressive notification strategy in consumer tech, your 'just add AI' feature strategy needs to be rooted in genuine user pull, not distribution leverage.

Tokenmaxxing Is Dead — What Replaces It?

Meanwhile, the practice of 'tokenmaxxing' — maximizing AI token consumption as a proxy for adoption — is dying. Meta employees and others adopted this metric as proof of AI engagement, but multiple sources confirm the tide is turning. Companies are shifting from 'use more AI' to 'use AI smarter,' which fundamentally changes the value equation.

This mirrors cloud computing's maturation: the market initially rewarded raw consumption, then pivoted to rewarding efficiency (serverless, spot instances, reserved capacity). AI tools are entering the same curve, but on a compressed timeline. Products that help users accomplish more with fewer tokens will outperform products that encourage maximum consumption.

What Actually Drives AI EBIT Impact

McKinsey's 2025 survey provides the counterpoint to both failures: workflow redesign — not AI feature bolting — is the strongest contributor to EBIT impact from generative AI. Yet only a minority of organizations have fundamentally redesigned even part of their operations. The gap between knowing the value of redesign and doing it is where both product opportunities and organizational advantages live.

Anthropic's Claude Code validates this: built by self-taught programmer Boris Cherny (not a PhD researcher pushing model benchmarks), it's driving Anthropic's revenue past $2.5B by understanding developer workflows, not maximizing model capability exposure. The product competition is shifting from 'what can the model do?' to 'what does the user need?'

The Revenue Model Implication

If your pricing is usage-based (per token, per API call, per query), model these scenarios now:

The companies that build value-based pricing alternatives now will have them ready when the efficiency shift accelerates.

Three Attack Vectors, One Strategic Problem

Anti-AI sentiment crossed from online anger into coordinated physical violence in Q1 2026, across three distinct and independently motivated threat vectors:

• Ideological: A 20-year-old allegedly threw a Molotov cocktail at Sam Altman's San Francisco home while his family slept. This follows a November 2025 murder threat that locked down OpenAI's offices.
• Local/NIMBY: Indianapolis councilman Ron Gibson's home was shot at 13 times with a 'NO DATA CENTERS' note, because he supports a datacenter project in Martindale-Brightwood.
• Geopolitical: Iran's Revolutionary Guard released satellite footage of OpenAI's Stargate campus in Abu Dhabi and promised its 'complete and utter annihilation.'

When politicians face bullets for supporting AI infrastructure, the regulatory environment shifts toward restriction — not enablement. Plan accordingly.

The Radicalization Pathway Matters for Timing

Current attackers are tied primarily to AI safety ideological movements, not displaced workers. This is an important nuance: the violence is currently contained to a specific radicalization pathway. But AI industry leaders are fueling it — Altman and Amodei publicly discussing white-collar job elimination without credible transition plans is described as making them 'look like psychopaths.' When economic displacement hits at scale, the pool of motivated backlash grows by orders of magnitude.

What This Means for Your Product

The implications are both immediate and structural:

Positioning & Messaging

The 'AI-powered' branding premium is eroding. Audit all customer-facing copy for displacement language — replace 'replaces,' 'eliminates,' 'automates away' with augmentation framing: 'helps you,' 'accelerates,' 'handles the tedious parts.' This isn't just optics; it should inform your actual product design. Human-in-the-loop workflows, explainability features, and visible attribution of AI assistance are competitive moats against a market increasingly hostile to AI that feels like it's replacing people.

Infrastructure Risk

The datacenter violence connects directly to legislative momentum already underway: Maine's LD 307 would pause datacenter construction through November 2027, and Michigan residents are organizing against energy costs. Politicians who face physical threats for supporting AI infrastructure will stop supporting it. If your compute depends on continued capacity expansion, model for delays.

The Vertical AI Advantage

OpenAI's Sora video tool dying while Netflix acquired Ben Affleck's vertical AI startup for up to $600M reinforces a parallel pattern: generic AI capabilities that scream 'look, AI!' are commoditizing and attracting backlash. Vertical AI products with deep domain integration — framed around the outcome, not the technology — capture value and avoid the branding liability. Lead with the outcome; let the AI be invisible infrastructure.