AI Usage at a Glance
Timeline
Jan 12, 2016
ModerationPractice documented: AutoMod uses machine learning and natural language processing to screen chat messages in real time, flagging potentially inappropriate content for human moderator review before it appears in chat.
Practice DocumentedView practice →May 18, 2017
ModerationNew evidence: AutoMod 2.0: Personalization Update
Evidence AddedView practice →Sep 26, 2018
RecommendationPractice documented: Twitch uses machine learning to personalize which live streams appear on each viewer's homepage, recommended channels sidebar, and mobile Discovery Feed, processing billions of requests daily to match viewers with streams based on viewing history and engagement patterns.
Practice DocumentedView practice →Aug 4, 2020
RecommendationPractice documented: Twitch rebuilt its search system from scratch starting in 2019, replacing a third-party dependency with an in-house system that uses Amazon OpenSearch for retrieval and machine learning models to re-rank search results for relevance.
Practice DocumentedView practice →Sep 14, 2020
Data AnalysisPractice documented: Twitch uses algorithmic ad-serving systems — integrated with Amazon's advertising platform since 2020 — to evaluate ad opportunities, match ads to audiences based on interest and behavior signals, and implement brand safety and fraud prevention measures.
Practice DocumentedView practice →Mar 9, 2021
Data AnalysisNew evidence: Twitch might be testing a streamer scoring system to facilitate ad sales
Evidence AddedView practice →Nov 30, 2021
ModerationPractice documented: Twitch's Suspicious User Detection uses machine learning to analyze account signals and chat behavior to identify users who create new accounts to circumvent channel-level bans, flagging them as "likely" or "possible" evaders for moderator review.
Practice DocumentedView practice →Dec 2, 2021
ModerationNew evidence: Twitch’s new tool uses machine learning to spot users evading bans
Evidence AddedView practice →Dec 15, 2021
ModerationNew evidence: Troll Recognition Twitch uses AI to flag trolls who try to avoid bans.
Evidence AddedView practice →Feb 10, 2022
ModerationPractice documented: Twitch uses a machine learning system at account signup to detect and prevent the creation of usernames containing offensive, hateful, or otherwise inappropriate content before accounts are created.
Practice DocumentedView practice →Mar 7, 2023
ModerationPractice documented: Following a high-profile deepfake scandal in January 2023, Twitch established explicit policies banning synthetic non-consensual exploitative images (NCEI) and updated its automated enforcement systems to detect and remove such content, with first-offense indefinite suspension.
Practice DocumentedView practice →May 7, 2023
ModerationNew evidence: Twitch now has a tougher stance against broadcasting deepfake porn
Evidence AddedView practice →Sep 28, 2023
RecommendationNew evidence: Twitch State of Engineering 2023
Evidence AddedView practice →Sep 28, 2023
Data AnalysisPractice documented: Twitch uses machine learning models trained on billions of daily data events to detect fraud, personalize gift subscription recipients, and optimize Hype Train engagement settings for each channel.
Practice DocumentedView practice →Sep 28, 2023
ModerationPractice documented: Twitch uses a machine learning model combining image analysis and text processing to automatically review custom emote submissions, instantly approving compliant emotes and flagging potential Community Guidelines violations for human review.
Practice DocumentedView practice →Sep 28, 2023
Data AnalysisNew evidence: Twitch State of Engineering 2023
Evidence AddedView practice →Sep 28, 2023
RecommendationNew evidence: Twitch State of Engineering 2023
Evidence AddedView practice →Oct 20, 2023
ModerationPractice documented: Smart Detection is an advanced AutoMod feature that uses deep learning models to learn individual channels' moderation preferences from moderator actions, providing personalized, scalable chat safety across over 95% of Twitch chat globally.
Practice DocumentedView practice →Jun 22, 2024
ModerationNew evidence: Smarter, Better, Faster: Using Machine Learning to Review Emotes
Evidence AddedView practice →Jul 29, 2024
RecommendationNew evidence: Find content you love faster with the new Twitch mobile app
Evidence AddedView practice →Aug 13, 2024
ModerationPractice documented: Twitch tested an ML-powered prompt system that detects potentially harmful chat messages before they are sent and asks users "Are you sure you want to send this?" — nudging them to reconsider without blocking the message outright.
Practice DocumentedView practice →Dec 13, 2024
ProductivityPractice documented: Twitch built a Slack-based AI chat assistant using Amazon Bedrock with retrieval-augmented generation (RAG) and agentic workflows to help Amazon ad sales teams instantly find Twitch advertising information, replacing a process that previously took up to 2 hours per query.
Practice DocumentedView practice →Mar 17, 2025
ModerationNew evidence: Machine Learning Summit: Twitch Chat Safety: Scalable and Personalized Moderation with Deep Learning
Evidence AddedView practice →Recent activity
Following a high-profile deepfake scandal in January 2023, Twitch established explicit policies banning synthetic non-consensual exploitative images (NCEI) and updated its automated enforcement systems to detect and remove such content, with first-offense indefinite suspension.
In January 2023, streamer Atrioc accidentally revealed deepfake pornographic images of fellow streamers during a broadcast, prompting widespread outrage. Twitch responded on March 7, 2023 with updated Community Guidelines: intentionally promoting, creating, or sharing synthetic NCEI results in indefinite suspension on the first offense. witch uses the term "synthetic NCEI" rather than "deepfake" to emphasize the non-consensual nature. The company consulted with the Cyber Civil Rights Initiative and UK Revenge Porn Helpline. In H2 2024, Twitch's transparency report shows automated and proactive detection systems working in tandem with user reports across all content categories, including nudity and exploitation.
Twitch uses machine learning models trained on billions of daily data events to detect fraud, personalize gift subscription recipients, and optimize Hype Train engagement settings for each channel.
The State of Engineering 2023 blog states: "Twitch generates billions of data events every day and our teams have specialized in building data pipelines and machine learning models to leverage that information. These teams empower a more intelligent, personalized and secure environment. Their work enables Twitch to select more interested gift subscription recipients, configure the most engaging Hype Train settings and identify instances of fraud." Twitch's data platform ("Tahoe") contains over 100 petabytes of data and ingests approximately 3 million events per second via the Spade event system. The Community Insights team processes over 10 million events per minute for streamer analytics.
Twitch uses algorithmic ad-serving systems — integrated with Amazon's advertising platform since 2020 — to evaluate ad opportunities, match ads to audiences based on interest and behavior signals, and implement brand safety and fraud prevention measures.
Per the State of Engineering 2023 blog, Twitch's Ad Product Engineering team builds "a robust set of services that evaluate opportunities to show ads and identify the best set and types of ads to show," alongside "brand safety, ad fraud prevention, and measurement tools." In 2020, Amazon opened Twitch inventory to programmatic buyers through Amazon Advertising, enabling interest-based targeting powered by Amazon's audience data. Twitch's proprietary SureStream technology (introduced 2016) stitches video ads directly into broadcast streams, enabling personalization "according to audience interest and product preference." Twitch also explored a Brand Safety Score system (discovered in internal API code in 2021) that automatically rates streamers on multiple factors for advertiser matching; Twitch confirmed the concept but stated "nothing has launched yet."
All practices by category
Following a high-profile deepfake scandal in January 2023, Twitch established explicit policies banning synthetic non-consensual exploitative images (NCEI) and updated its automated enforcement systems to detect and remove such content, with first-offense indefinite suspension.
Twitch tested an ML-powered prompt system that detects potentially harmful chat messages before they are sent and asks users "Are you sure you want to send this?" — nudging them to reconsider without blocking the message outright.
Twitch deployed Harmful Chatter Detection, a machine learning feature rolled out in August 2024 that flags both potential ban evaders and potential harassers in a channel before they post. The feature extends the earlier Suspicious User Detection tool and surfaces flagged users to channel moderators for review.
Twitch uses a machine learning system at account signup to detect and prevent the creation of usernames containing offensive, hateful, or otherwise inappropriate content before accounts are created.
Twitch's Suspicious User Detection uses machine learning to analyze account signals and chat behavior to identify users who create new accounts to circumvent channel-level bans, flagging them as "likely" or "possible" evaders for moderator review.
Twitch uses a machine learning model combining image analysis and text processing to automatically review custom emote submissions, instantly approving compliant emotes and flagging potential Community Guidelines violations for human review.
Smart Detection is an advanced AutoMod feature that uses deep learning models to learn individual channels' moderation preferences from moderator actions, providing personalized, scalable chat safety across over 95% of Twitch chat globally.
AutoMod uses machine learning and natural language processing to screen chat messages in real time, flagging potentially inappropriate content for human moderator review before it appears in chat.
Twitch uses machine learning models trained on billions of daily data events to detect fraud, personalize gift subscription recipients, and optimize Hype Train engagement settings for each channel.
Twitch uses algorithmic ad-serving systems — integrated with Amazon's advertising platform since 2020 — to evaluate ad opportunities, match ads to audiences based on interest and behavior signals, and implement brand safety and fraud prevention measures.
Twitch rebuilt its search system from scratch starting in 2019, replacing a third-party dependency with an in-house system that uses Amazon OpenSearch for retrieval and machine learning models to re-rank search results for relevance.
Twitch uses machine learning to personalize which live streams appear on each viewer's homepage, recommended channels sidebar, and mobile Discovery Feed, processing billions of requests daily to match viewers with streams based on viewing history and engagement patterns.
Related Companies
Trace Foundation
iZotope
Krafton
ElevenLabs
Activision Blizzard
Have evidence about Twitch's AI practices? Submit a report.
Submit a report →AI Trace is free and nonprofit. Support our work