safetywashing
In plain wordsMaking AI work sound as if it improves safety when it mostly just makes AI more capable — like 'greenwashing' for environmental claims.
The practice, or the accusation, of presenting AI work as safety progress when it mainly advances general capabilities, or of using safety language and commitments for reputational benefit without corresponding risk reduction. The term covers marketing claims, research framed as alignment that chiefly improves product usefulness, and voluntary pledges without verification. A 2024 study from the Center for AI Safety gave it an empirical form: many published AI safety benchmarksbenchmark, n.: A fixed, shared test — a set of questions or tasks with an agreed way of scoring — used to compare AI systems with one another and to track progress over time. correlated strongly with general capability scores across models, so gains on them could reflect scale rather than safety-specific progress; the authors recommended that new safety benchmarks report this correlation. Accusations of safetywashing are themselves disputed, since some safety-relevant properties, such as robustnessrobustness, n.: An AI system's ability to keep working correctly and safely when conditions change — unfamiliar data, noisy inputs, or people deliberately trying to trick it. or truthfulness, are expected to improve with capability.
Calling a better math score an alignment result is safetywashing if the benchmark just tracks capability.
Origin
Formed on the model of 'greenwashing' and 'ethics washing'. An early use is Adam Scholl's LessWrong post 'Safetywashing' (July 2022); Richard Ren, Steven Basart, Dan Hendrycks and colleagues later gave the term an empirical definition in 'Safetywashing: Do AI Safety Benchmarks Actually Measure Safety Progress?' (2024).
Usage note
Compare AI washingAI washing, n.: Falsely claiming that a product or company uses AI, or exaggerating how much it does, to look more advanced and attract customers or investors., which exaggerates the use of AI itself. The label is a criticism, so its application to any particular organization is a matter of dispute.
Related
- AI washingAI washing, n.: Falsely claiming that a product or company uses AI, or exaggerating how much it does, to look more advanced and attract customers or investors.
- AI hypeAI hype, n.: Exaggerated claims about what AI can do or will soon do, made to attract money, attention or customers, which can mislead the public and policymakers.
- benchmarkbenchmark, n.: A fixed, shared test — a set of questions or tasks with an agreed way of scoring — used to compare AI systems with one another and to track progress over time.
- alignment taxalignment tax, n.: The price of caution in AI: whatever performance, money or time a developer sacrifices to make a system safe and aligned, compared with an otherwise identical system built with no such care.
- responsible scaling policyresponsible scaling policy, n.: A public rulebook in which an AI company promises to test each new model for dangerous abilities and to add stronger safeguards, or slow down, as its models become more capable.
- AI safetyAI safety, n.: The effort to keep AI from doing serious harm, whether because it pursues aims its makers never intended, is turned to bad uses or simply malfunctions — a concern that grows as systems become more capable.
Sources
- Scholl, A. (2022). Safetywashing. LessWrong.
- Ren, R., Basart, S., Khoja, A. et al. (2024). Safetywashing: Do AI Safety Benchmarks Actually Measure Safety Progress? arXiv:2407.21792. arxiv.org/abs/2407.21792