Malayalam Wikipedia's AI bot made 888 edits, none reverted, and its checks blocked 380

A volunteer-built AI bot called AksharaBot has made 888 edits to Malayalam Wikipedia without a single revert so far, according to a post on Wikimedia's Diff blog on October 12. Its safeguards stopped another 380 proposed changes, and it has listed about 7,660 factual problems for human editors instead of fixing them.

The bot is allowed to repair language only. Facts, numbers, dates, names and references are off limits, even when it can see they are wrong.

The problem it was built for

Many Malayalam articles were machine-translated from other languages, with odd results. In one cricket article, the surname of Everton Weekes was rendered as the Malayalam word for weeks, according to the post by Netha Hussain, who proposed and runs the bot.

Hussain brought the idea to the Malayalam community in August 2026. Because the AI's rewrites change from run to run, a traditional code review could not prove it was safe. Instead the bot first edited 100 articles with every change published, then results from 700 articles in three stages, then a supervised trial on 100 more. Approval came by the end of that month.

Three checks that must all agree

Each edit passes through three steps. First, one AI agent suggests a correction, with no power to save anything. Ordinary code then compares the before and after text and throws the edit out if any link, template, reference or digit changed, including Malayalam numerals. Finally, a second AI agent, which never sees the first one's reasoning, judges whether only the wording changed or the meaning did too.

The checks are strict enough to block good edits. The bot once proposed changing a urine test result in a medical article from "white" to "red blood cells". Because that altered a fact rather than the language, it was refused and sent to humans.

Why the cheaper models were dropped

Hussain uses Anthropic's Claude, naming the Fable and Opus models, and runs each article through two or three passes. Cheaper models were rejected. "I tested cheaper models, and they returned sentences that read fluently but had quietly lost part of their meaning," Hussain wrote, calling it the riskiest failure because the output gives no sign of a problem.

The factual problems the bot finds, such as outdated officeholders, infoboxes that contradict their articles and sports pages frozen in 2020, go to public tables with tick-boxes. A task force of Malayalam Wikipedians works through them.

The code is open for reuse, though Malayalam is the only edition it has run on so far. The post invites other language editions with poorly translated backlogs to build similar tools; how the safeguards hold up in another language is still untested.