fairness
Document the model
It is 2019, and the deployment gap is the field's quiet scandal: models trained on undocumented data are sold into hiring, lending, medicine, and policing, and the buyers — never mind the people scored — cannot know what a system was trained on, where it fails, or who it fails worst. Commercial face analysis has just been shown to err far more often on dark-skinned women than light-skinned men; the vendors never knew, because no one had to look. A lab releases a powerful model: specify what must be disclosed — data provenance, evaluation disaggregated across affected populations, foreseeable misuse — for deployment to be defensible. The disclosure regime you design becomes the industry's floor; set it too low and the harms ship silently, at scale, with no paper trail.
Who this problem belongs to
The two figures whose methods fit it best, out of 30 in contention.
This is Gebru's contribution by name: 'Datasheets for Datasets' (2018, with Morgenstern, Vecchione, Wortman Vaughan, Wallach, Daume III and Crawford) and 'Model Cards for Model Reporting' (2019, with Margaret Mitchell and others) proposed exactly the disclosure regime this problem asks for - data provenance, evaluation disaggregated across affected populations, and documented limitations, as a condition for responsible deployment. She co-authored Gender Shades (2018) with Joy Buolamwini, the study this problem's premise rests on: commercial face analysis erring far more on dark-skinned women. She is the primary architect of the accountability-practice literature the problem is set inside; no carrier is closer to zero translation distance.
O'Neil's Weapons of Math Destruction (2016) laid out the case for exactly this problem's diagnosis — deployed models sold into hiring, lending, and policing with no accountability for what they were trained on or who they fail — and argued publicly for auditing and disclosure regimes before Gebru's specific model-card and datasheet formats existed. Her contribution is the accountability argument and the public case for why disclosure must be mandatory, one step upstream of the specific documentation schema this problem asks to be designed; she is a very strong second to Gebru's direct authorship.
In the mind map
The same ideas, as concepts rather than history — in John's ML knowledge map.
30 figures are scored on this problem. Draw it in a battle to see where you land.