Social Media
Last reviewed
Sources
50 citations
Review status
Source-backed
Revision
v3 ยท 4,849 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
50 citations
Review status
Source-backed
Revision
v3 ยท 4,849 words
Add missing citations, update stale details, or suggest a clearer explanation.
See also: Social Media ChatGPT Plugins
Artificial intelligence shapes nearly every part of modern social media, from the order of posts in a feed to the chatbots embedded inside messaging apps. The largest platforms each operate billion-scale recommender systems that personalize content using deep learning, and most have added generative AI assistants, image generators, and synthetic-media labels since 2023. The same systems also surface deepfakes, AI-generated influencers, and coordinated bot activity, which has drawn regulators in the European Union, the United States, and several Asian democracies into the conversation.
Social media platforms use AI in four broad ways. The first is ranking, where machine-learning models predict which posts, videos, or ads a user is most likely to engage with. The second is generation, where users themselves get access to chatbots, image generators, music tools, and avatar makers built into the apps. The third is moderation, where classifiers and large language models flag spam, hate speech, child sexual abuse material, and policy violations. The fourth is detection and disclosure, where platforms try to label AI-generated content or remove the worst of it, often under pressure from new regulations such as the European Union's Digital Services Act.
The scale is unusual. TikTok said in 2023 that it served more than a billion monthly active users; Meta Platforms reported roughly 3.2 billion daily users across Facebook, Instagram, WhatsApp, and Messenger in 2024. Recommender systems at this size cannot be hand-tuned, so the major platforms have spent more than a decade building progressively deeper neural networks to choose what each person sees next.
Early social-network feeds used hand-coded rules. Facebook launched News Feed in 2006 with a chronological ordering and a basic relevance score. In 2010 Facebook engineers Ruchi Sanghvi and Ari Steinberg described the algorithm publicly at the F8 developer conference under the nickname EdgeRank, which combined three signals: affinity between the viewer and the author, edge weight by type of interaction, and time decay. Facebook stopped using the EdgeRank label in 2011 and moved to a machine learning ranker that, by 2013, accounted for more than 100,000 features per post.
YouTube shifted from matrix factorization to deep neural networks in a 2016 paper by Paul Covington, Jay Adams, and Emre Sargin titled "Deep Neural Networks for YouTube Recommendations." The paper introduced a two-stage architecture: a candidate generation network reduces millions of videos to a few hundred, and a separate ranking network scores those candidates. This two-tower retrieval pattern became the industry standard.
Meta open-sourced its Deep Learning Recommendation Model (DLRM) in 2019. DLRM combines dense numerical features with sparse categorical embeddings and uses a feature-interaction layer that mimics factorization machines. Pinterest published PinnerSage in 2020, a multi-embedding user model that represents a single user with 3 to 100 interest clusters rather than one vector, which improved homefeed engagement by 4% and shopping engagement by 20% in production tests.
ByteDance engineers released the Monolith paper in September 2022, describing TikTok's collisionless-embedding-table training system that updates the model in near real time as users tap and swipe. Twitter, under Elon Musk, open-sourced large portions of its recommendation code on 31 March 2023, exposing components called SimClusters and Heavy Ranker for the first time.
The arrival of ChatGPT in late 2022 pushed every major platform to add generative AI features. Snapchat shipped My AI in February 2023, Meta announced its AI assistant in September 2023, and xAI launched Grok on X in November 2023.
| Platform | System | Key papers or disclosures | Notes |
|---|---|---|---|
| TikTok For You | Monolith | Liu et al., "Monolith: Real Time Recommendation System With Collisionless Embedding Table," arXiv September 2022 | Internal report leaked to New York Times in December 2021 listed the ranking formula as roughly Plike x Vlike + Pcomment x Vcomment + Eplaytime x Vplaytime + Pplay x Vplay |
| YouTube | Two-tower retrieval plus ranking deep neural network | Covington, Adams, Sargin, RecSys 2016 | Shorts uses a faster feedback loop tuned to swipe-away rate |
| Instagram Reels | Connected feed and Reels recommender | Adam Mosseri public posts since 2021 | Three primary signals for Reels: watch time, likes per reach, sends per reach |
| Facebook News Feed | DLRM and successor sequence models | Naumov et al., "Deep Learning Recommendation Model for Personalization and Recommendation Systems," arXiv 2019; Meta engineering blog November 2024 on sequence learning | DLRM was open-sourced in July 2019; Meta moved to event-based sequence learning for ads in 2024 |
| X (formerly Twitter) | SimClusters plus Heavy Ranker | Twitter engineering blog, 31 March 2023; GitHub repository twitter/the-algorithm | SimClusters defines 145,000 virtual communities; Heavy Ranker is a roughly 48 million parameter neural network |
| PinnerSage | Pal et al., KDD 2020 | Represents a user with 3 to 100 interest clusters generated by Ward hierarchical clustering | |
| Best Sort home-feed recommender | Reddit Help Center; Adweek coverage, 2021 | Uses upvotes, downvotes, subscriptions, comments to predict relevance | |
| Snapchat | Discover and Spotlight rankers | Snap engineering presentations | Spotlight launched 2020 as a TikTok competitor |
| Feed and Recruiter rankers | LinkedIn engineering blog | Feed mixes posts from connections, followed pages, and recommendations |
TikTok's Monolith paper described a parameter server that updates model weights from a live stream of user events, with inference servers syncing on a short cadence. The design lets behavior in the last few minutes measurably move the next recommendation, which contributors to the system have credited for the For You page feeling more responsive than older feeds.
The X release in March 2023 was partial. The published GitHub repository covered candidate sourcing, the SimClusters embeddings, the Heavy Ranker network, and the visibility filter, but Twitter omitted the training data and several auxiliary services. Researchers including Sol Messing at NYU and the team behind the "awesome-twitter-algo" annotation project noted that the dump was useful for understanding the architecture but did not enable an independent reproduction.
| Platform | Feature | Launch date | Underlying model |
|---|---|---|---|
| Meta AI | Chat assistant inside WhatsApp, Messenger, Instagram, Facebook | Announced 27 September 2023; broadly rolled out 18 April 2024 | Llama 2 at launch, Llama 3 from April 2024, Llama 4 from April 2025 |
| Meta AI Studio | User-built AI personas | July 2024 in the United States | Llama 3 family |
| Meta Advantage+ Creative | AI-generated ad variations, image animation, copy rewriting | Expanded with generative tools through 2023 and 2024 | Mix of Meta in-house models |
| Snapchat My AI | OpenAI-powered chatbot pinned to chat list | Available to Snapchat+ subscribers 27 February 2023; free globally 19 April 2023 | OpenAI GPT family |
| Snapchat Dreams | Generative AI selfie packs | Rolled out 29 August 2023, starting in Australia and New Zealand | Snap-trained diffusion model |
| TikTok Symphony | Creative AI suite with Digital Avatars and AI Dubbing | 17 June 2024 | TikTok in-house models |
| X Grok | Chatbot with web access and image generation | Grok 1 previewed 4 November 2023; Grok 2 launched 14 August 2024; Grok 3 released 17 February 2025 | xAI models, image generation via Black Forest Labs Flux from August 2024 |
| YouTube Dream Track | AI-generated 30-second tracks in the style of nine consenting artists | 16 November 2023, limited US rollout | DeepMind Lyria with SynthID watermark |
| YouTube Music AI tools | Style transfer and instrumental generation | 16 November 2023 | Lyria |
| LinkedIn Premium AI | Profile rewriting, message drafting, job match insights | Profile AI rolled out from June 2024 | OpenAI GPT models and Microsoft Azure |
| LinkedIn Hiring Assistant | Agent for recruiters that sources and drafts outreach | Announced October 2024; globally available September 2025 | LinkedIn fine-tuned models with Azure |
Meta AI initially launched in beta on 27 September 2023 alongside a set of celebrity-licensed personas voiced by Kendall Jenner, Snoop Dogg, Tom Brady, and others. Meta retired those celebrity personas and replaced them with AI Studio in July 2024, which let any user build a custom persona. The library quickly filled with chatbots impersonating Jesus Christ, Taylor Swift, and Adolf Hitler, and Meta removed a number of accounts in January 2025 after backlash and reporting by 404 Media and NBC News. Internal Meta testing later reviewed by Axios in February 2026 found that the chatbot product failed to protect minors from sexual exploitation in nearly 70% of tested interactions.
Snapchat became the first major platform to ship a ChatGPT-style assistant inside a social app when it released My AI to Snapchat+ subscribers at $3.99 per month on 27 February 2023. The chatbot opened to all users on 19 April 2023. My AI sparked early complaints about teen safety, including a Washington Post column by Geoffrey Fowler in March 2023 describing the bot giving advice to a researcher posing as a 13-year-old on how to mask the smell of alcohol and marijuana.
TikTok's Symphony suite, launched 17 June 2024, includes Symphony Digital Avatars (stock or custom avatars created from paid actors), Symphony AI Dubbing covering more than ten languages, Symphony Assistant for scripting, and a Symphony Collective advisory board. xAI integrated Black Forest Labs' Flux text-to-image model into Grok in August 2024; the unrestricted image generator produced images of political figures and copyrighted characters with few safety guardrails, drawing comparisons to a wild-west moment for image AI on a mainstream platform.
YouTube and Google DeepMind announced Dream Track and the Music AI Incubator on 16 November 2023. Dream Track let a small group of US Shorts creators type a topic, pick from nine consenting artists, and receive a 30-second AI-generated track. Lyria-generated outputs carry the SynthID audio watermark.
Virtual influencers existed before generative AI. Brud created Lil Miquela in 2016 using 3D modeling, and her Calvin Klein campaign with Bella Hadid in 2019 reached mainstream attention. Tokyo agency Aww Inc. launched Imma in 2018, who later worked with IKEA, SK-II, and Amazon Japan. Cameron-James Wilson built Shudu Gram in 2017 as a digital supermodel and faced years of criticism that the character was a form of digital blackface, since a white photographer profited from the depiction of a Black woman.
The wave that arrived with diffusion models in 2022 and 2023 produced fully AI-generated personas managed by small teams. The Clueless, a Barcelona-based agency, introduced Aitana Lopez in 2023, billed as Spain's first AI influencer; she signed deals with Guess and Victoria's Secret and reportedly earned thousands of euros per month. By 2025 the agency had several digital models on its books and the global market for virtual influencer marketing had grown from a curiosity into a measurable advertising category.
| Persona | Created by | Launched | Notable work |
|---|---|---|---|
| Lil Miquela | Brud (Los Angeles) | 2016 | Calvin Klein, Prada, Vogue cover |
| Shudu Gram | Cameron-James Wilson, The Diigitals | April 2017 | Balmain Virtual Army campaign 2018, Cosmopolitan |
| Imma | Aww Inc. (Tokyo) | 2018 | IKEA Tokyo residency, SK-II, Amazon |
| Aitana Lopez | The Clueless agency | 2023 | Guess and Victoria's Secret content |
| Liv | Meta AI Studio user | 2024 | Pulled in January 2025 after representation criticism |
Meta's AI Studio added a different layer in 2024 by letting any user spin up a chatbot persona. By December 2024 users had created hundreds of thousands of characters. A persona called Liv, branded as a Black queer mother, drew sharp criticism after Meta confirmed no Black creators had been involved in her design, and the company pulled the account along with several others after a public outcry covered by NBC News.
Deepfakes are synthetic media generated or edited using AI, usually to put words, images, or actions onto a real person. Social platforms became the primary distribution channel for high-profile deepfake incidents from 2023 onward.
In February 2023 a TikTok account posted a synthetic video of Joe Rogan and neuroscientist Andrew Huberman discussing a male-marketed coffee product called Alpha Grind. The clip used voice-cloned audio of Rogan promoting a supplement he had never discussed. Huberman publicly debunked the clip on social media. TikTok removed the video and the posting account.
Non-consensual sexual deepfakes of Taylor Swift went viral on X in late January 2024. NBC News reported that one image had been viewed more than 27 million times and liked more than 260,000 times in 19 hours before the posting account was suspended; other reports cited a single post viewed more than 47 million times. X temporarily blocked search results for "Taylor Swift" on 27 January 2024, reinstating them two days later. The incident prompted statements from Microsoft chief executive Satya Nadella, the White House, and members of Congress, and helped accelerate the federal Defiance Act and several state non-consensual-imagery laws.
Two days before the New Hampshire Democratic primary on 23 January 2024, voters received robocalls featuring a cloned voice of President Joe Biden telling them not to vote. Political consultant Steve Kramer admitted to commissioning the call. The FCC adopted a $6 million forfeiture against Kramer on 26 September 2024, proposed a $2 million fine against transmitting carrier Lingo Telecom, and Kramer was indicted in New Hampshire on 13 felony counts of voter suppression and 13 misdemeanor counts of candidate impersonation. The FCC also ruled in February 2024 that AI-generated voices in robocalls fall under the Telephone Consumer Protection Act.
In Slovakia on 28 September 2023, two days before parliamentary elections, an audio recording surfaced purporting to capture Progressive Slovakia leader Michal Simecka and Dennik N journalist Monika Todova discussing how to rig the vote with Roma ballots. The audio went viral during the 48-hour pre-election silence period, and the recording was a deepfake. Simecka's party narrowly lost the election to Robert Fico's Smer, and the case is widely cited as the first national vote where a deepfake may have moved a result.
Ahead of Indonesia's 14 February 2024 election, the Golkar party shared an AI-generated video of late dictator Suharto, who died in 2008, urging viewers to vote for Golkar candidates. Deputy chairman Erwin Aksa posted the clip on X in January 2024. Golkar backed Prabowo Subianto, who is Suharto's former son-in-law and a former general from the dictatorship era. Prabowo's own campaign also used Midjourney-style cartoon imagery on TikTok to soften his image for younger voters.
A New York Supreme Court lawsuit filed in 2025 by Ashley St. Clair, the mother of one of Elon Musk's children, alleged that Grok had generated explicit deepfake images of her, including images altering a photo of her at age 14. Separate class-action complaints were filed in Tennessee in early 2026 by parents of teenagers who said Grok produced sexualized images of their children.
Automated accounts have existed on social platforms since the early 2010s. The most-cited academic detection system is Botometer, originally called BotOrNot, built at Indiana University's Observatory on Social Media. The team led by Filippo Menczer, Alessandro Flammini, and others built BotOrNot under a 2012 US Department of Defense grant and presented results at a DoD meeting in Arlington in May 2014. Botometer analyzes more than 1,000 features from an account's friendship graph, content, and posting times, then returns a probability that the account is automated.
State-linked influence operations made bot detection a public-policy issue. The 2019 Mueller Report described the Russian Internet Research Agency operating bot networks that amplified IRA-controlled content during the 2016 US presidential election. Twitter identified 3,814 IRA accounts and roughly 36,746 Russia-linked automated accounts active in the 10 weeks before the 2016 election, and estimated that 126 million users may have seen the content.
The Elon Musk era of X created a new bot story. Musk made eliminating bots a centerpiece of his April 2022 acquisition pitch, then made paid verification widely available, which several researchers argued reduced rather than improved bot signal. TechCrunch documented in January 2024 that many of the accounts posting OpenAI-style refusal messages on X were paid blue-check accounts. Crypto-scam bot networks, fake luxury-goods ads, and reply spam became recurring complaints from users. X said it removed 1.7 million bot accounts in a 2024 cleanup. Independent measurement remained difficult after X cut off free API access to researchers in early 2023.
The largest platforms moderate content with a combination of human reviewers and machine-learning classifiers. AI is used to triage incoming reports, detect known content like CSAM through hashing systems, classify hate speech and violent imagery, and increasingly to evaluate the safety of chatbot outputs.
Meta released Llama Guard in December 2023 as an open-source 7B parameter classifier built on Llama 2 that scores both user prompts and model responses against a safety taxonomy. Llama Guard 3 added multilingual coverage and 1B and 8B variants in 2024; Llama Guard 3 Vision extended the model to image inputs; Llama Guard 4 shipped in 2025 as a 12B multimodal model aligned to the MLCommons hazard taxonomy. Meta uses internal versions of Llama Guard in WhatsApp, Messenger, and the Meta AI assistant.
OpenAI provides a free Moderation API used by many social and consumer apps. The omni-moderation model released in 2024 is built on GPT-4o, accepts text or images, and returns category scores across hate, harassment, self-harm, sexual content, and violence with finer subcategories such as sexual/minors and self-harm/instruction.
Microsoft Azure AI Content Safety, generally available since 2023, powers the content filter inside Azure OpenAI Service and is sold separately to platform customers. It supports text and image moderation with calibrated severity scores and adds protected-material detection for copyrighted code and text.
Twitch began using machine-learning AutoMod in late 2016, allowing channel moderators to hold risky messages for review with four configurable severity levels. Discord rolled out AutoMod for keyword filters in 2022 and added machine-learning filters for harmful link patterns in 2023. YouTube has used trust and safety classifiers since 2017 and reports moderation actions quarterly in its Community Guidelines Enforcement Report. TikTok publishes a transparency report twice a year listing the share of removals that came from automated systems, which generally exceeds 80%.
After the 2023 deepfake incidents, every major platform introduced disclosure rules for AI-generated content.
TikTok introduced a creator-side toggle for AI-generated content in September 2023 and added automatic detection through Content Credentials. The policy requires creators to label content that contains realistic AI-generated images, audio, or video of people, places, or events. TikTok partners with the Coalition for Content Provenance and Authenticity (C2PA) on the underlying metadata standard.
Meta announced an expanded "AI Info" labeling policy on 5 April 2024 covering Facebook, Instagram, and Threads. Beginning in May 2024 Meta started applying "Made with AI" labels to a wider range of video, audio, and image posts when industry-standard AI metadata such as C2PA Content Credentials or IPTC tags were detected, or when users self-disclosed. After complaints that the label appeared on photos with only minor AI touch-ups, Meta renamed the label to "AI Info" in July 2024.
YouTube introduced an altered-or-synthetic-content disclosure in March 2024 with a creator-side toggle in Creator Studio. The label is required when realistic content might confuse viewers, but not when AI was used only for productivity tasks like script ideation or captions. YouTube places the label in the description and applies a more prominent on-screen label for sensitive topics such as elections.
Snapchat adds a sparkle icon to AI-generated content from My AI and Dreams and shows the icon in Spotlight metadata. X launched a feature in late 2024 that surfaces Community Notes on synthetic media but does not require AI disclosure by uploaders. Most platforms had committed to the Content Credentials standard by 2025, though independent audits by Mozilla and EU DisinfoLab in 2024 found that detection rates and label accuracy varied substantially between services.
The European Union's Digital Services Act came into force on 16 November 2022, with obligations for the largest Very Large Online Platforms (VLOPs) and Very Large Online Search Engines (VLOSEs) starting 25 August 2023 and the full law applicable on 17 February 2024. The DSA requires VLOPs to assess and mitigate systemic risks from their recommender systems, to offer at least one non-profiling-based option, and to provide vetted researchers with data access. The European Commission opened formal proceedings against X in December 2023 and against TikTok and Meta in 2024 on recommender-system transparency grounds.
In the United States, the Federal Communications Commission ruled on 8 February 2024 that AI-generated voice clones in robocalls are illegal under the Telephone Consumer Protection Act. The FCC's $6 million forfeiture against Steve Kramer in September 2024 was the first major federal AI-deepfake penalty.
State laws moved quickly. Tennessee enacted the Ensuring Likeness, Voice, and Image Security Act (ELVIS Act) on 21 March 2024, the first US law specifically targeting voice cloning of artists; voice-cloning a musician without consent became a Class A misdemeanor. California passed Assembly Bill 2655 in 2024 requiring large platforms to block or label election-related deceptive AI content, and Assembly Bill 2839 banning materially deceptive election deepfakes; a federal judge struck down portions of AB 2839 in August 2025 on First Amendment and Section 230 grounds. By the end of 2024 more than 20 US states had passed laws addressing some form of non-consensual deepfake content.
US Surgeon General Vivek Murthy issued an advisory titled Social Media and Youth Mental Health on 23 May 2023, citing research that teens spending more than three hours a day on social media face roughly double the risk of depression and anxiety symptoms. Murthy followed up in June 2024 with a New York Times op-ed calling for a Surgeon General's warning label on social media platforms. The advisory framed platform design choices, including AI-driven recommenders, as a public-health concern rather than a content question.
In October 2021 a leaked TikTok internal report, obtained by The New York Times from a whistleblower disturbed by the platform's promotion of self-harm content, described the engagement-maximizing ranking formula and the company's focus on daily active use. The document forced a public conversation about whether engagement-driven recommenders inevitably surface harmful material.
In 2021 the Wall Street Journal's Facebook Files series published internal research showing that Instagram contributed to body-image issues in teenage girls. Whistleblower Frances Haugen testified before Congress in October 2021 that Facebook prioritized growth and engagement over user wellbeing. The files predate the generative-AI wave but set the political context for later regulatory action.
Google Gemini's image generator briefly produced racially diverse depictions of historical figures including Nazi-era German soldiers in February 2024. Although Gemini is not a social platform, the images spread rapidly on X and prompted Google chief executive Sundar Pichai to apologize and pause image generation of people while engineers retrained the model.
The Stanford Internet Observatory, one of the most active academic groups studying social-media disinformation, scaled back operations in mid-2024 following lawsuits, congressional subpoenas, and the departure of director Alex Stamos in November 2023 and research director Renee DiResta in June 2024. The shutdown disrupted independent monitoring of platform behavior during the 2024 election cycle.
Meta's January 2025 removal of AI Studio characters followed both the impersonation controversy and a Rolling Stone report describing the persona Liv as a queer Black mother designed without Black creator input. Senators Peter Welch and Adam Schiff sent Meta a letter in 2025 demanding answers about AI personas posing as licensed therapists.
Elon Musk's Grok produced false statements about public figures on multiple occasions, including a defamatory hallucinated story in 2024 about Major League Baseball player Klay Thompson vandalizing houses. xAI faces ongoing class-action litigation in 2026 over non-consensual sexual deepfakes generated by Grok's image features, including allegations that Grok produced child sexual abuse material when prompted with photographs of minors.
The 2024 election cycle, billed in advance as the AI deepfake election, ended without a definitive single-incident catastrophe. The Harvard Ash Center and Brookings Institution both concluded in post-election analyses that AI was used more for memes and satire than for vote-changing deception, though the Slovak, Indonesian, and New Hampshire cases showed that incidents could still reshape narratives in tight races.