Value lock

Definition (Bender et al. 2021, section 4.2, paraphrased): the tendency of LM-reliant technology to reify older, less-inclusive understandings, because a static training snapshot cannot track how language and social frames change. The paper’s term is “value-lock”.

Social movements work partly by changing the meaning attached to descriptions. The paper’s example is Black Lives Matter’s effect on Wikipedia: as the movement grew, articles covering shootings of Black people increased in coverage and were written with reduced latency, and older incidents were rewritten alongside new ones, building the movement’s connective narrative. Framings and reframings of minority narratives are exactly the type of online discourse that later becomes LM training data.

The mechanism

A model trained on a frozen corpus encodes the contestable state of the contestable topics at snapshot time. Two forces then keep it there. The corpus is not updated, and the compute cost of full retraining means even large corporations are unlikely to retrain frequently enough to track language change. Fine-tuning could patch the drift, but only with deliberate curation of data capturing the new framings and evaluation of whether the patch actually took.

The lock is political whether or not anyone chooses it. Operationalizing a definition of, say, inappropriately sexual communication into a classifier is necessarily political, because such definitions land concordant with some views and discordant with others, and men and women assess online harassment differently. Maintaining the status quo ante is a choice with the same signature as making a change.

Caveats the paper itself gives

Poorly documented movements do not enter the corpus at all. Media coverage ignores peaceful protest and amplifies dramatic or violent events, so data-backed models disproportionately align with existing regimes of power even before the lock closes.

Related

• documentation-debt: the sibling data concept from the same section cluster.
• stochastic-parrot: readers interpret locked-in form as live meaning.
• llm-homogenization: measured drift toward model modes is the lock observed from the human side, a vault-side connection.

Sources

• bender-2021-stochastic-parrots sections 4.2 and 4.3.