I Tried to Make AI Writing Sound Human by Banning AI Words Through Logit_bias

5 hours ago

Codex CLI now displays Markdown tables responsively in the terminal, improving readability in narrow windows.
The author tested logit_bias (an API setting that changes token selection scores) to reduce AI-typical words in generated text.
A blacklist of words common in AI writing was converted to token IDs and given negative bias values to lower their selection probability.
Prompt-based blacklists offer flexibility but can be ignored or cause awkward substitutions, while logit_bias applies direct pressure without understanding context.
Experiments with DeepSeek V4 Flash showed that logit_bias removed about one-third of watched words but occasionally damaged meaning or grammar.
Compatibility for logit_bias varies: OpenAI's newer models reject it, Anthropic lacks a native parameter, but OpenRouter supports it for many models with provider-specific caveats.
Hard bias values like -100 can remove words but cause repetition loops, while moderate values like -8 offer a trade-off between word reduction and output quality.
The author concludes that neither prompt blacklists nor logit_bias fully solve the issue of making AI writing sound human; each has distinct failure modes.

Invest in upfront alignment by making implicit assumptions explicit in the initial prompt to avoid unexpected AI outputs.
Restart with better context rather than steering AI when initial outputs are far off; avoid compounding errors through path dependence.
Equip AI agents with the same tools you use (CLI, MCP, browser control) to handle setup, testing, and instrumentation tasks.

The author discovered that lazy evaluation of research papers works: by writing without citations, readers will send references.
Bayesian spam filtering was known since 1998, but early filters (Pantel & Lin) had poor performance (92% catch, 1.16% false positives), while the author's later filter achieved 99.5% catch with <0.03% false positives.
Five key differences explain the performance gap: training on too little data, ignoring message headers, stemming tokens, using all tokens instead of the 15 most significant, and not biasing against false positives.

The author's code writing has shifted from 90% human-written to 90% AI-written in eight months.
The task involved optimizing a matrix-multiply kernel using Cluster Launch Control (CLC) on GPUs, where AI assisted in finding, understanding, fixing, and benchmarking code.
AI helped identify unused shared memory to fix a deadlock, and suggested multiple optimizations, leading to faster code for problematic inputs.

A conversational AI with long-term memory stores (vector and fact) had an extractor that over-produced junk milestones (~6.5/day), causing prompt monoculture with forced romantic declarations.
Diagnosis used a trace inspector and golden set; a keyword-as-necessary-condition rule blocked 94% of historical junk while passing 100% of positive controls.
Fix A stopped the tap: keyword gate, roleplay guard, tightened dictionaries, anti-multi-label, deployed behind a test suite.

AI visibility tools promise to measure brand visibility in AI answers but often provide false precision by presenting tidy claims like mention rates and rankings.
Scraping the frontend of AI products like ChatGPT or Claude captures only one synthetic session with many uncontrolled variables, leading to biased measurements.
Even with identical prompts, AI systems can produce varying answers due to factors like model batching, personalization, and nondeterministic behavior.

Invest in upfront alignment by making implicit assumptions explicit in the initial prompt to avoid unexpected AI outputs.
Restart with better context rather than steering AI when initial outputs are far off; avoid compounding errors through path dependence.
Equip AI agents with the same tools you use (CLI, MCP, browser control) to handle setup, testing, and instrumentation tasks.

The author discovered that lazy evaluation of research papers works: by writing without citations, readers will send references.
Bayesian spam filtering was known since 1998, but early filters (Pantel & Lin) had poor performance (92% catch, 1.16% false positives), while the author's later filter achieved 99.5% catch with <0.03% false positives.
Five key differences explain the performance gap: training on too little data, ignoring message headers, stemming tokens, using all tokens instead of the 15 most significant, and not biasing against false positives.

The author's code writing has shifted from 90% human-written to 90% AI-written in eight months.
The task involved optimizing a matrix-multiply kernel using Cluster Launch Control (CLC) on GPUs, where AI assisted in finding, understanding, fixing, and benchmarking code.
AI helped identify unused shared memory to fix a deadlock, and suggested multiple optimizations, leading to faster code for problematic inputs.

A conversational AI with long-term memory stores (vector and fact) had an extractor that over-produced junk milestones (~6.5/day), causing prompt monoculture with forced romantic declarations.
Diagnosis used a trace inspector and golden set; a keyword-as-necessary-condition rule blocked 94% of historical junk while passing 100% of positive controls.
Fix A stopped the tap: keyword gate, roleplay guard, tightened dictionaries, anti-multi-label, deployed behind a test suite.

AI visibility tools promise to measure brand visibility in AI answers but often provide false precision by presenting tidy claims like mention rates and rankings.
Scraping the frontend of AI products like ChatGPT or Claude captures only one synthetic session with many uncontrolled variables, leading to biased measurements.
Even with identical prompts, AI systems can produce varying answers due to factors like model batching, personalization, and nondeterministic behavior.

Hasty Briefsbeta