Accessibility features are exactly the kind of unglamorous, high-value work that's easy to skip under deadline pressure, which makes it a good test case for what AI assistance is actually worth versus what's still marketing.
The genuinely automatable wins
Generating subtitle tracks from dialogue audio, and checking a color palette against common forms of color blindness to flag risky combinations before they ship, are both things current tools handle reliably and cheaply. Neither used to be free, in time or attention, and now both are close to it. There's very little excuse left to skip them.
Where it helps but doesn't finish the job
Suggesting alternative control remapping options, or flagging UI text that's too small against a target resolution, are things AI-assisted tooling can catch and propose fixes for, but the actual decision about what remapping options make sense for your specific game still needs a human who understands the controls.
Where it doesn't really help yet
Difficulty accessibility, options for players with different reaction time or motor control needs, is a design problem, not a generation problem. It requires understanding which parts of your specific game are the hard part for which kind of player, and building real alternate paths through it. No current tool automates that judgment, because it's the same "is this actually good" problem that shows up everywhere else AI-assisted design still needs a human.
Why this is worth building into a generation pipeline from the start
Retrofitting accessibility after a game is finished is expensive and often incomplete, subtitle systems bolted on late tend to miss lines, remapping added as an afterthought tends to have gaps. A pipeline that generates subtitle tracks and checks color palettes as a normal part of every build, not a separate pass someone has to remember to run, is far more likely to actually ship with these features intact.
The honest scope
AI assistance meaningfully lowers the cost of the easy, mechanical accessibility wins. It does nothing to lower the cost of the harder design questions, and treating the easy wins as "we did accessibility" is how a game ships with subtitles and colorblind mode and still isn't actually accessible to the players who needed the harder parts.