Ai = Bad

Why AI Is a Bad Deal: The Critics’ Case

This post is generated with ai…

In a randomized controlled trial, 16 experienced open-source developers took 19% longer on 246 real tasks when allowed to use AI. They believed they had been sped up by 20%.
The gap between felt speed and measured speed is the real finding.
Source: METR (2025), Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/

Across 211M changed lines (2020 to 2024): refactored (“moved”) lines dropped from ~25% to under 10%, copy/pasted lines rose from 8.3% to 12.3%, and duplicated blocks rose roughly 8x in 2024.
2024 was the first year on record where copy/paste exceeded refactoring.
Duplication is not just ugly: every future fix has to be applied in N places.
Sources: GitClear, AI Copilot Code Quality (2025) https://www.gitclear.com/ai_assistant_code_quality_2025_research and The Maintainability Gap (2026) https://www.gitclear.com/the_ai_code_quality_maintainability_gap

Users with an AI assistant wrote significantly less secure code than the control group, and were more likely to believe their code was secure.
Overconfidence is the dangerous part, not the bug itself.
Source: Perry, Srivastava, Kumar & Boneh, ACM CCS 2023 https://arxiv.org/abs/2211.03622

EEG study with 54 participants: the LLM group showed the weakest, least distributed brain connectivity; the unaided group the strongest.
Most LLM users could not quote a sentence from the essay they had just “written.”
The authors call this cognitive debt: cheap now, expensive later in recall, ownership and critical thinking.
Source: Kosmyna et al., MIT Media Lab (2025) https://arxiv.org/abs/2506.08872
Caveat: preprint, small sample, one task type, one model.

If the easy tasks are automated, the path that produced senior engineers disappears.
You cannot review what you never learned to write.
This is an argument from structure, not from a study. Treat it as a hypothesis, not evidence.

Free assistants are not free. The product is the conversation log.
A US court ordered OpenAI in May 2025 to preserve output logs that users had deleted; the order was lifted in October 2025, but data from that window remains stored. Later, 20M de-identified logs were ordered handed to news plaintiffs.
Lesson: “delete” is a UI affordance, not a physical fact. Your prompts are discoverable.
Sources: Engadget (Oct 2025) https://www.engadget.com/ai/openai-no-longer-has-to-preserve-all-of-its-chatgpt-data-with-some-exceptions-192422093.html; Zuboff, The Age of Surveillance Capitalism (2019)

OpenAI began showing ads to US Free and Go users on 9 February 2026 and opened a self-serve Ads Manager later that year, reportedly hitting a $1B annualized run rate in under 200 days.
Targeting uses conversation topics and chat history, not just keywords.
OpenAI says ads never influence the answer. The structural worry is that once revenue depends on attention, the incentive to keep you talking outlives the promise.
Sources: OpenAI, Testing ads in ChatGPT https://openai.com/index/testing-ads-in-chatgpt/; explainx.ai (2026) https://explainx.ai/blog/openai-advertise-in-chatgpt-ads-launch-july-2026

Training data is scraped without consent or payment (see NYT v. OpenAI, filed Dec 2023).
Content moderation and data labelling is outsourced to low-wage workers.
Source: Perrigo, TIME (2023), on Kenyan workers paid under $2/hour https://time.com/6247678/openai-chatgpt-kenya-workers/

Inference at scale, not just training, is the growing share of energy use; generative tasks are far more expensive per query than discriminative ones.
Source: Luccioni, Jernite & Strubell, Power Hungry Processing, ACM FAccT 2024 https://arxiv.org/abs/2311.16863

A handful of firms control the models, the compute and the distribution.
Critique of the underlying economics: Bender, Gebru, McMillan-Major & Mitchell, On the Dangers of Stochastic Parrots, FAccT 2021 https://dl.acm.org/doi/10.1145/3442188.3445922

Even at temperature 0, LLMs are not guaranteed to give the same output for the same input. Floating-point arithmetic is non-associative, GPU kernel execution order depends on batch size and concurrent load, and tiny rounding differences can flip which token wins when two candidates are nearly tied — that single flip then cascades through the rest of the autoregressive output.
A structured benchmark sending the same prompt 10 times at temperature 0 found determinism collapsing sharply on longer, open-ended prompts across major models.
Providers acknowledge this openly: OpenAI calls its outputs only “mostly deterministic” even with a fixed seed; Anthropic’s own docs say results won’t be fully deterministic at temp=0 and recommends sampling multiple times to cross-validate.
Consequence: the same prompt against the same model can yield a different bug fix, a different security judgment, a different citation, or a different answer to the same question — on different days, different hardware, or under different load. That undermines debugging, scientific replication, and auditing.
Sources: QAnswer, Why LLMs Are Not Deterministic Even at Temperature 0 (2026) https://www.qanswer.ai/blog/llm-non-determinism-temperature-zero; He et al., Nondeterminism in LLM Inference https://arxiv.org/pdf/2506.09501; LLMs Get Lost In Multi-Turn Conversation https://arxiv.org/pdf/2505.06120

METR now labels its own result historical and published newer data on late-2025 tools in February 2026. One RCT, 16 developers, tools from early 2025.
GitClear measures correlation. Code volume rose too; some duplication may reflect more shipping, not worse judgement.
Google’s DORA 2024 report found AI adoption correlated with a ~3.4% improvement in code quality alongside a ~7.2% drop in delivery stability. The picture is mixed, not uniformly bad.
The MIT study is a preprint with 54 participants and the authors themselves ask people not to read it as “brain damage.”
Most of these findings are about how the tools are used, not about whether they can be used well.