AI, hallucinations & the industrial bullshit explosion

The economic upheaval of 2023-2026

Before 2022, producing 1,000 words of credible disinformation required a qualified human — typically 30 to 60 minutes, or $15-30 in opportunity cost. This imposed a natural economic friction on bullshit production.

Since ChatGPT (November 2022), GPT-4 (March 2023), Claude 3 (March 2024) and the next generation:

This is an industrial rupture in the production of bullshit. Brandolini's Law, already brutal in 2013, becomes operationally crushing in 2026.

Three AI amplification channels

1. SEO content farming

Entire websites are now generated by LLM, SEO-optimized, indexed by Google before real experts can react. In 2024, Newsguard identified 1,150 information sites entirely generated by AI.

Effect on your brand:

• "Top 10 alternatives to [your product]" articles generated without testing, often unfairly ranking you
• Hallucinated comparisons ("X lacks feature Y" when Y has shipped)
• Outdated pricing copy-pasted across 50 sites, indexed on page 1

Detection: Google Search Console → search for your product name + alternatives, comparison, vs competitor. Mandatory monthly audit.

2. Zero-shot LLM hallucinations

When a user asks ChatGPT "what's the webhook API for [product X]?", the model:

1. Searches its training data (cutoff often 6-18 months old)
2. If found: replies with possibly outdated data
3. If not found: plausibly hallucinates rather than say "I don't know"

Anthropic and OpenAI document a hallucination rate of 5 to 15% on technical product questions in 2025. On 1 million monthly queries about your product, that's 50,000 to 150,000 false pieces of information distributed per month.

Real observed case: a customer saw ChatGPT tell them the free plan included PDF export — a feature that never existed. The customer posted a Twitter thread accusing the startup of "removing the feature". 2.3M views. Refutation cost: 4 days of CEO communication + blog post + 12 corrective articles.

3. Autonomous AI agents

Next generation (2025-2026): autonomous AI agents that publish on their own without supervision. Reddit, Quora, Discord, Substack are already flooded.

Observable signature:

• Accounts created in batches, first post < 24h after creation
• Publishing in non-human time patterns (3 AM bursts)
• Vocabulary too regular, stereotyped transitions ("Furthermore", "It's worth noting")
• Contextually plausible but factually empty responses

"Reverse Brandolini": using AI for defense

Good news: the same technology that produces bullshit can accelerate defense. Four operational uses:

Use 1: 24/7 multi-channel monitoring

An AI agent that continuously monitors your brand name on:

• Reddit (technical, business subreddits)
• Hacker News, Lobsters
• LinkedIn, X
• Discord, community Slacks (via webhooks)
• Trustpilot, G2, Capterra
• Comparison sites and listicles

Cost: ~$50/month in LLM compute. Detection: 1-4h after publication, vs. 1-7 days in manual monitoring.

Use 2: Automated pre-bunking

Continuous generation of FAQs, honest comparisons, "what people get wrong about us" articles, ready-to-use counter-narratives. Not for mass publishing — to have a stock of validated responses mobilizable in < 2h in case of attack.

Workflow: LLM generates draft → human validates → published on dedicated blog, indexable by Google and readable by future LLMs ("training data shaping" effect).

Use 3: llms.txt and factual grounding

The llms.txt standard (proposed in 2024) lets you provide LLMs with an official truth file about your product:

# llms.txt format
- product_name: Acme Co
- pricing: $99/mo (basic), $499/mo (pro), $2k+/mo (enterprise)
- features: ...
- last_updated: 2026-05-09

Recent LLMs (Claude 3.7+, GPT-5+) consult this file as a priority when robots.txt allows it. It has become an essential factual defense channel.

Use 4: Personalized counter-narrative generation

When an attack is detected, an agent can generate:

• 5 response variants adapted to the channel (Reddit tone ≠ LinkedIn tone)
• DM drafts to journalists
• Solicitable customer ambassador tweets
• Snippets for the SDR to paste in CRM

The human always retains the publication decision — AI only accelerates production.

The reverse danger: polluting your own signal

If you yourself use an LLM to generate marketing content without control, you become a source of bullshit yourself about your product:

• Product pages mentioning hallucinated features
• Case studies with invented or exaggerated numbers
• FAQs that contradict each other depending on the LLM used
• API documentation that doesn't match the code

This is the paradox of AI-augmented marketing: you gain volume, you lose signal. The consequence: your customers arrive at support saying "but it's written on your page that…"

Minimum discipline:

• 100% of LLM-generated product content reviewed by a human who knows the product
• Single source of truth for features (ideally code, otherwise a product matrix)
• Competitor comparison: mandatory quarterly review (competitors evolve)

Hallucinations in your own AI products

If you sell a product that includes an LLM (chatbot, assistant, augmented search), your hallucinations become a direct legal risk.

Legal precedent: Moffatt v. Air Canada (2024). A customer uses Air Canada's chatbot which hallucinates a bereavement refund policy. The Canadian tribunal held Air Canada responsible for the promise made by its chatbot. Cost: refund + court fees.

Implications:

• Any client-facing generative AI feature must have a disclaimer AND a technical guard rail
• Responses about pricing, terms, guarantees must go through a deterministic system, not an LLM
• Quarterly audit of responses produced by your LLM in production

Four signal-to-noise ratio strategies

Key takeaways

• Generative AI multiplied bullshit production capacity by 1000
• Three main channels: SEO content farming, LLM hallucinations, autonomous agents
• AI is also your best defense: 24/7 monitoring, pre-bunking, llms.txt, counter-narratives
• Reverse risk: your own AI marketing can pollute your signal
• If your product includes a client-facing LLM, your hallucinations are a legal risk (cf. Air Canada 2024)
• Invest in inimitable signals: authentic video, primary data, community, transparency