CHICustomer Hostility Index
← Back to Special Investigations

CHI Special Investigation · Two-Regime Comparison

X / TwitterWhat Actually Changed

Twitter built the machine.X changed what it was for.

Twitter already controlled reach, accounts, developers and customer remedies with sweeping discretion. Under X, those powers became faster, more personal, more deeply monetised and — in one important case — were turned on customers who stopped buying.

The machinery Largely inherited

Twitter wrote the discretion clause, built the visibility tooling, squeezed its developers, sold distribution and ran non-refundable paid terms. X did not need to invent any of it.

The operating philosophy Changed

Who exercised the power, how fast, for whose benefit, what it cost, and what recourse remained. Discretion became personal, rapid, monetised, written into contract and adversarial toward counterparties who resisted.

No Twitter antecedent of comparable kind Customer-to-Adversary Conversion

Advertisers whose remedy had been to stop buying became defendants in an antitrust suit. High confidence for advertisers; not found for subscribers, creators or developers.

2016Earliest capture of the “limit distribution or visibility” clause. Word-for-word unchanged through the 10 Apr 2026 X terms
5Monetisation components classified as new in kind under X. Everything else intensified a Twitter antecedent
4Terms-of-service revisions narrowing user remedies, Sep 2023 – Apr 2026
0Twitter-era suits against an advertiser located in the reviewed record

This is not a verdict on Elon Musk, on Twitter's politics or on X's. It is a comparison of two regimes of one institution: Twitter, Inc. to the close of the acquisition on 27 October 2022, and X from that date to the research cutoff of 2 September 2026. No CHI score is issued on this page.

● 20 minute read · Research cutoff 2 September 2026
01

The part everyone gets wrong

There was no clean baseline.

The popular story has a before and an after: a platform that served its users until a new owner bought it. The primary record does not support the before. Most of the principal powers X used over users, developers and customers already existed at Twitter.

2018Restricted visibilityDown-ranked replies and search results from accounts showing “behavioral signals”, with no author notice described.
2011Controlled developersTold developers the answer to “should I build a client?” was “no”, then capped the ones that had.
2015Monetised dependencyTerminated firehose resellers and moved enterprise data into its own priced channel.
FY2021Sold distributionPromoted Products placed paid posts in timelines, search and replies: about 89% of revenue.
2016Account discretionReserved the right to terminate accounts “for any or no reason”.
2021Non-refundable paid serviceTwitter Blue's terms: “All payments … are final and not refundable.”
2009–22Privileged verified statusLegacy-verified accounts carried maximum reputation weight in Twitter-era ranking code.
Wrong reading one Twitter was customer-friendly until Musk bought it.

Contradicted by the primary text: the same discretion clause, the same visibility code library, the same decade-long developer squeeze. The largest penalty for a user-data harm against either regime in the record belongs to Twitter.

Wrong reading two Nothing changed except the owner.

Also contradicted. The remedy architecture, the sale of status, litigation against customers who stopped buying and the concentration of discretion in one person have no equivalent in the Twitter record.

The powers mostly existed before Musk. What changed was who exercised them, how quickly, against whom, for whose benefit, what became monetised, and what recourse customers kept.

02

Regime A · Twitter, to 27 October 2022

Twitter built the levers.

Eight pre-acquisition antecedents, each dated and primary-sourced where the record allows. Each card closes with what X later did with the same lever.

The visibility clause

The Terms of Service let the company limit distribution or visibility of any Content … without liability to you.

Kept word for word through every X version examined, to 10 Apr 2026.

Termination for “any or no reason”

Twitter could suspend or terminate an account, or stop providing all or part of the service, at any time for any or no reason.

Reworded 29 Sep 2023 to “no reason at our convenience”. More explicit, not broader.

The developer squeeze

“Should I build a client?” — “no” (11 Mar 2011). A 100,000-user-token cap on client apps (Aug 2012). Streaming removed in 2018, with the replacement priced at $2,899 a month per 250 users.

Third-party clients cut off without notice in Jan 2023; free API access ended Feb–Mar 2023.

Access as a competitive weapon

Instagram's find-friends access cut after the Facebook acquisition (Jul 2012). Meerkat's social-graph access cut on 13 Mar 2015 with about two hours' notice, in the days around the Periscope announcement. Data reseller DataSift said it was “blindsided”.

Rival and publisher links blocked, stripped of engagement or delayed, Dec 2022 – Aug 2023.

Undisclosed reach reduction

Down-ranking by “behavioral signals”, no author notice described. Two months later: “We do not shadow ban,” under a definition that excluded the practice. Account-level “Do Not Amplify” and blacklist tags surfaced in the Twitter Files (a curated release with no denominator data).

Restricted-reach labels made rule-based limits visible to the author (Apr 2023).

Paid distribution

Promoted Products: labelled, auction-priced placement in timelines, search and conversations, about 89% of revenue ($4.52bn of $5.08bn). Follower Ads sold labelled account prominence.

Payment entered the ranking of organic posts and replies, without a label.

Twitter Blue

The earliest located terms (effective 1 Sep 2021) were non-refundable, auto-renewing, discontinuable “with or without notice” and bound to arbitration with a class waiver. The price rose 67% for existing subscribers in 2022; Edit was gated behind Blue.

Reissued almost verbatim on 9 Nov 2022, then scaled to status, reach and earnings.

Status with privileges

Verification was editorial, and removable “at any time without notice” as a conduct sanction (Nov 2017). Twitter-era ranking code gave legacy-verified accounts maximum reputation mass and a search-relevance feature. Free, and not for sale.

The status itself was put on sale.
$150m

Twitter was not benign. The largest monetary penalty for a user-data harm against either regime in the record is Twitter's: the FTC/DOJ order of 25 May 2022 for using phone numbers and email addresses that more than 140 million users had supplied for security, to target advertising between 2014 and 2019. Severity-weighted, it remains the single largest hidden degradation in this investigation.

03

Regime B · X, 27 October 2022 – 2 September 2026

X changed the operating system.

Twitter exercised discretion too. The finding is not that X used discretion, but how. Twitter's was institutional and executive-led: a small senior group working through written policy, an advisory council and published rationales. X's was the owner, in public, often within the day.

Same lever
How Twitter used it
How X used it
Removing and restoring accountsAccount discretion
Twitter

— @realDonaldTrump permanently suspended with a same-day published rationale citing two posts, a prior warning and a written policy.

X

— Trump reinstated after a poll. A “general amnesty” followed a second poll days later. No written criteria were located for either.

Making the rulesRule creation
Twitter

Written policies, an advisory Trust & Safety Council (est. 2016) and biannual transparency reports in a stable format.

X

— Council dissolved by email. A “live location” rule created on the 14th; about ten journalists who had reported on the episode suspended on the 15th, most restored two days later after a poll.

Announcing policyOwner posts as policy
Twitter

No located instance of comparable directness. Executive statements described policy; they were not the policy.

X

A word designated a slur by owner post (21 Jun 2023). A creator-payout exclusion “deleted” by post (17 Jul 2023). Demonetisation of Community-Noted posts announced by post (29 Oct 2023). Link deprioritisation confirmed by post (24 Nov 2024).

Cutting off developersEcosystem control
Twitter

— Streaming shutdown announced about eight months ahead, then delayed two months after a developer campaign.

X

— Third-party clients cut off without notice; “long-standing API rules” cited without naming one; a clause banning substitute services added two days later.

Ranking the person in chargeOwner self-preference
Twitter

No equivalent located.

X

— Platformer reported that engineers built a multiplier boosting the owner's posts roughly 1,000-fold after his Super Bowl post under-performed. Reported · Medium

Backing downReversal under pressure
Twitter

— The NY Post laptop-story block reversed to labelling within about 48 hours; the CEO called it “wrong”.

X

— A ban on links to rival platforms withdrawn within about a day; a delay on rival and news links reversed the same afternoon. Fast reversals are a genuine X trait.

PersonalFrom a senior group applying written policy to one owner deciding in public.
RapidRules written and enforced inside a day. Price windows of about 30 days.
MonetisedStatus, security, reach and earnings eligibility attached to payment.
ContractualisedFour terms revisions narrowing remedies; a tariff declared “not a penalty”.
AdversarialCounterparties who resisted met with litigation.

Twitter's institutions dampened, delayed and sometimes disclosed the same powers that X exercised openly, quickly and personally.

04

Forced Monetization

Everything became a tier.

Twitter charged for features and sold placement. X attached payment to things Twitter had given away: status, a security control, the right to earn, the right to read, and access to the data itself. Five components are new in kind. The count is deliberately narrow: everything else X charged for intensified something Twitter already ran.

1
Twitter · VerificationFree, editorially granted status

Granted for thirteen years on Twitter's judgement. Its 2020–21 policy made promoting the sale of verification badges a ground for ineligibility.

X · 9 Nov 2022 / 20 Apr 2023Purchasable status

$7.99 a month at launch; relaunched 12 Dec 2022 at $8 web / $11 iOS. Legacy badges removed on 20 Apr 2023 with no transition or appeal path located. Organisations: $1,000 a month.

2
Twitter · SMS two-factorA free security control

Text-message two-factor authentication available to any account.

X · 20 Mar 2023Subscribers only

SMS 2FA restricted to Blue subscribers. Authenticator apps stayed free. The stated reason was SMS fraud costing about $60m a year; keeping SMS for payers points to cost-driven gating rather than a security retirement.

3
Twitter · Creator programmesThresholds, not a toll

Tips (no cut) and Super Follows (10,000 followers; 3%, then 20%): numeric thresholds and disclosed splits. No creator paid Twitter to qualify.

X · 13 Jul 2023Pay to be eligible to earn

Ads revenue sharing required a paid subscription plus impression thresholds. From 8 Nov 2024 payouts were computed on engagement from paying users, under a formula X did not disclose.

4
Twitter · ReadingLimits were for developers

Rate limits applied to API clients. No located limit on how much a person could read.

X · 1 Jul 2023Read limits by payment status

Daily caps of 600 posts for unverified accounts and 6,000 for verified (by then largely paid) accounts, raised the same day to 1,000 and 10,000. Whether they persist is unresolved; 2025 code exempts Premium viewers from a rate-limit gate. Medium on the Jul 2023 figures

5
Twitter · ScrapingProhibited, not priced

The terms barred scraping (2018–22) without attaching a price.

X · 15 Nov 2024A private tariff on access

$15,000 per 1,000,000 posts requested, viewed or accessed in any 24 hours, declared “not a penalty”. It bites only above a million posts a day, which excludes ordinary users and most academic projects.

Intensified, not invented

Twitter antecedent; X changed price, speed or scope

  • Feature gating. Twitter gated Edit and long video behind Blue; X extended it to post length, Articles and more.
  • Repricing. Blue rose 67% with about three months' notice (2022). Premium+ went $16 → $22 → $40 in about sixteen months, with windows of about 30 days.
  • Developer access. Twitter's 2012 caps and 2018 pricing, then X's end of free read access and a $100–$42,000+ monthly ladder (2023).
  • Advertiser eligibility. From 21 Apr 2023, buy a checkmark or spend $1,000 a month to run ads (fees reportedly waived for the ~10,000 largest advertisers).
  • Government checks. The grey check, introduced without a stated fee, priced at $1,000 a month from 7 Oct 2025.
Changed in beneficiary

Same mechanism, different recipient

  • Status-linked ranking. Twitter's code gave prominence to editorially chosen accounts; X gave it to paying ones. Covered in the next section.
  • Reply prioritisation. Twitter sold labelled placement; X sold an unlabelled preference in replies, graduated by price tier from Oct 2023.
  • Creator payouts. Pay scales with engagement from paying viewers, making creators dependent on X's subscription sales.
Not extraction

Where the paid tier added something

  • Creator cash. X paid ad revenue to ordinary individual creators, which no Twitter programme did (about $20m by Sep 2023, per X).
  • Grok. A new product, later opened to free users with limits (Dec 2024).
  • ID verification. A new, optional check for paid users outside the EU/UK (Sep 2023).
  • A free tier for posting survived throughout.
05

Pay-for-Visibility

Paying for visibility.

Visibility was for sale under both regimes. What changed is what you bought and whether anyone was told. Twitter sold labelled advertising and gave status-linked prominence away. X attached prominence to a status anyone could buy, and did not label the effect.

Twitter

Labelled ads. Free status in the ranking code.

  • Promoted Products placed paid posts in timelines, search and replies, marked “Promoted” and priced by auction. They did not change how an account's organic posts ranked.
  • Follower Ads sold labelled account prominence.
  • Legacy verification fed the ranking substrate: Twitter-era code assigned verified accounts maximum reputation mass and carried a verified-author search feature. High that it entered ranking Medium on magnitude
X

Purchasable status in organic ranking.

  • Paid reply prioritisation. A stated “slight preference” for subscribers' replies from Dec 2022; a “reply boost” graduated by price tier from 27 Oct 2023. Still described on X's Help Center. High
  • Verified-first ranking code published in 2023 (below). High that the mechanism existed
  • Deception finding. The European Commission found that the purchasable checkmark “deceives users” (preliminary Jul 2024; decision 5 Dec 2025, €120m).

A Blue-author scorer with default multipliers

VerifiedAuthorScalingScorer
  in-network   default 4.0
  out-of-network default 2.0
  live value: served from unpublished
  feature switches
Mechanism: existed

The open-sourced ranking code multiplied Blue-verified authors' For You scores by these defaults. It also partitioned candidates verified-first.

What users actually received

live multiplier values   not disclosed
date removed            not disclosed
verified-only For You   announced; not shown to ship
Unresolved

No located page told users the magnitude. A Nov 2022 study found no engagement boost from paid verification at that point.

The current For You ranker

xai-org/x-algorithm @ 85ac72a
  ranking scorer: no premium or
  verified term
  reply prioritisation: still stated
Paid For You boost: not in published code

Absence from published code does not establish absence from production: some files are unpublished. A paid For You boost in 2025–26 is unresolved. Low–Medium

How to read the 4.0 and 2.0 They are code defaults in the 2023 release, not measured reach. Live production values were controlled separately and were not disclosed. This page does not say, and the evidence does not show, that paying accounts received four times or twice the reach.
06

Customer-to-Adversary Conversion

When a customer becomes an adversary.

This is the one mechanism in the investigation with no Twitter-era antecedent of comparable kind. It is established at High confidence for one group: advertisers, the paying customers who walked away at scale.

  1. Advertisers pause

    General Motors pauses the day after the acquisition closes; VW, United, General Mills, Pfizer, Mondelez and others follow within weeks.

  2. “A thermonuclear name & shame”

    The owner threatens advertisers he says are pressured by “activist groups”. No such list was located.

  3. A second pause wave

    After the owner replies “You have said the actual truth” to an antisemitic conspiracy post and Media Matters publishes ads shown beside pro-Nazi content, IBM, Apple, Disney, Comcast and Lionsgate pause.

  4. “Go f*** yourself”

    On stage at DealBook, the owner tells departing advertisers this, names Disney's chief executive, and says “the whole world will know”.

  5. X sues its former customers

    A Sherman Act §1 group-boycott suit in the Northern District of Texas against the World Federation of Advertisers and its GARM initiative, Unilever, Mars, CVS and Ørsted. WFA closes GARM within days.

  6. Settle and return

    Unilever is dismissed after saying X had “committed to meeting our responsibility standards”.

  7. Seven more defendants

    Nestlé, Abbott, Colgate-Palmolive, Lego, Tyson, Shell and Pinterest are added (Twitch was reportedly added in Nov 2024).

  8. Threats in sales talks, reported and denied

    The Wall Street Journal reports that at least six companies struck ad deals after X threatened to add them to the suit. X's chief executive called it “salacious clickbait”. Medium · denied

  9. The court rules

    Judge Boyle (Dkt. 255) dismisses the domestic advertisers with prejudice for failure to plead an agreement to boycott or antitrust injury. As reported, the court described their conduct as consistent with independent decisions not to buy.

  10. Appeal and settlement

    X notices a Fifth Circuit appeal (Apr). X and WFA settle the surviving claim on 29 Jul, with WFA pledging not to restart GARM. An appellate brief against the advertiser defendants is reported filed in August. Medium on the brief

Domestic advertiser defendantsDismissed with prejudice

Failure to plead an agreement to boycott or antitrust injury. X appealed; the appeal was pending at the cutoff.

Shell International, Lego A/S, Nestlé S.A., ØrstedDismissed without prejudice

Lack of personal jurisdiction over these foreign entities, not a ruling on the merits.

World Federation of AdvertisersSurvived until settlement

WFA's motion to dismiss was denied. The claim was not adjudicated on the merits before it settled on 29 Jul 2026.

Adjacent evidence · not the mechanism

X also sued critics whose research informed advertisers, and suspended journalists who reported on the owner. These facts are established at High confidence, but these parties were platform users, not paying commercial counterparties. They corroborate a wider adversarial posture; they do not add to the advertiser finding.

CCDHSued 31 Jul 2023. Dismissed under California's anti-SLAPP law on 25 Mar 2024; Judge Breyer: “This case is about punishing the Defendants for their speech.” Ninth Circuit appeal undecided at the cutoff.
Media MattersSued in Texas on 20 Nov 2023, then in Ireland and Singapore. The Texas motion to dismiss was denied (29 Aug 2024) and the case was in discovery at the cutoff; Singapore's High Court dismissed X's appeal on service (19 Jun 2026).
Journalists and rivalsAbout ten journalists suspended under a same-day rule (Dec 2022). Links to rival platforms and news outlets blocked or delayed after public disputes. Medium for competitors and publishers

Where the mechanism was not found. No located case shows X treating an individual paying subscriber, a creator or a developer who disputed its terms as an adversary. For those groups the mechanism is unresolved or absent in the reviewed record, and this page does not extend it to them.

07

Remedy Asymmetry

The contract changed too.

This is not only a story about an owner's posts. X rewrote the paper. Twitter's user terms were thin but stable for six years. X narrowed them in four steps in 31 months, while adding remedies for itself.

Twitter baseline

California law, San Francisco courts. No arbitration, no class waiver, no limitation period in the main terms.

Class-action waiver

The first X-authored terms, pre-posted 23 Aug. Also an express scraping ban.

Texas, and a tariff

Texas law; an exclusive Tarrant County / N.D. Tex. forum for users while X may sue users where they live; the $15,000 tariff; an AI-training licence “with no compensation”; a one-year limitation period.

Retroactive forum

The forum clause applied “regardless of when the conduct … arose” and extended to affiliates. Limits split: one year federal, two years state.

A second county

Wichita County added as a forum. No advance notice banner was captured for this revision.

Binds users generallyTerms of Service
Twitter

California courts; no arbitration or class waiver.

X at cutoff

Class waiver; Texas forum with X's asymmetric right to sue at the user's home; one- or two-year limits. No consumer arbitration clause in any version to the cutoff.

Binds paying subscribersPurchaser Terms
Twitter · 1 Sep 2021

Non-refundable. AAA arbitration in San Francisco, class waiver, one-year limit, 30-day opt-out.

X · 1 Aug 2025

Arbitration dropped for exclusive Tarrant County courts; class waiver kept; two-year limit (EU/UK: Irish courts, one year). No refund if X suspends the account or withdraws features.

Binds ad-revenue creatorsCreator Revenue Sharing Terms
Twitter

No equivalent programme.

X · 1 Aug 2025

Payment calculation changeable “for any reason”; X may withhold, set off or recoup; termination “with or without cause”. AAA arbitration kept, with class waiver, one-year limit and opt-out.

Binds subscription creatorsSubscriptions Creator Terms
Twitter

Super Follows terms (2021) were not retrieved.

X · 1 Aug 2025

Texas courts, no arbitration, two-year limit. Creators should have “no expectation that any particular revenue percentage … will apply”.

What did not change

The discretion itself

The visibility clause. The $100 liability-cap amount (its drafting was widened in 2024 and 2026). Appeals in both regimes' terms point to a help-centre form, with no entitlement to reasons or a timeline. Paid services were non-refundable under both.

Refunds · bounded negatives

None located, which is not the same as none

No company-issued refund was located in the reviewed record for November 2022 checkmark buyers whose badges vanished, for Twitter Blue subscribers who lost ad-free articles on 31 Oct 2022 (under Twitter's own terms), or for subscribers to the third-party clients cut off in 2023. Apple's automatic pro-rated refunds for those apps fell 70–85% on the developers.

Where users gained

The EU is a different story

EU and UK users gained statutory redress under the DSA and Online Safety Act, and X published EU appeal reversal rates: 10.8% (Apr–Sep 2024), then 25.6% for suspensions and 28.7% for restricted reach in its Oct 2025 report. No Twitter-era equivalent was located. Every captured X terms revision except Apr 2026 was pre-posted 25–37 days ahead.

Confidence. High on the contract terms, which are primary documents. Which regime gave users the more effective remedy in practice is unresolved, and EU statutory remedies prevent a universal adverse finding. The intensification is contractual and US-centred.

08

Counter-evidence

What X did better.

X did not become more opaque in every dimension. On several measures it disclosed more than Twitter had, and on some it reversed bad decisions faster. These are genuine mitigants, and they bound every finding on this page.

Restricted-reach labels

Rule-based reach limits became visible to the author and to viewers, with a feedback channel. X's own announcement conceded that platforms, Twitter included, “have not historically been transparent” about such limits.

Feedback “does not guarantee” that reach is restored.

A written enforcement ladder

A public appeals form and a graduated-enforcement standard, published as policy rather than applied case by case.

An appeal still carries no entitlement to reasons or a timeline.

Published appeal reversals

EU transparency reports gave the first quantified appeal-reversal rates in either regime: 10.8%, then 25.6% for suspensions and 28.7% for restricted reach.

Produced under DSA compulsion, and EU-only.

Open-sourced ranking code

Two ranking codebases published; the Aug 2026 release added numeric weights, visibility-filtering and label code and a dated election filter. The 2023 release is how the Blue multiplier became known at all.

Some files remain unpublished; live configuration is not disclosed.

Community Notes

Expanded under X. The bridging design requires agreement across raters who usually disagree, and studies in the record find noted posts are reshared roughly half as much.

Studies also find notes often arrive after a post's viral phase.

Public-utility API exemption

Verified government and publicly owned services were exempted from API charges, and free scaled access for “Public Utility” apps was kept when pricing changed again in 2026.

It came after public-safety bots lost access and New York's MTA quit the platform.

“Under the Hood”

A pilot letting account holders download the labels applied to their own account: a per-account transparency tool with no Twitter-era equivalent located.

A pilot; its reach beyond the first cohort by the cutoff is unresolved.

ID verification

An optional government-ID check for subscribers, adding a real identity signal to a checkmark that had become purchasable.

Paid users only, and not offered in the EU/UK at launch.

Fast reversals

The rival-link ban lasted about a day, the Substack restriction about three, the news-link delay an afternoon, the login wall about five days. A creator-payout change was paused the day Nigerian creators objected (Mar 2026).

Reversals tracked public pressure more than any stated process.

The balance. The institutions X removed were mostly ones that constrained the company: the Trust & Safety Council, a Media Rating Council pre-audit, submissions to the Lumen database, the free academic API. The ones it added are mostly disclosures that do not constrain it, with EU regulation the exception. That does not make them trivial. Several are transparency Twitter did not offer, and a reader who skips them has misread the investigation.

09

Political asymmetry

Politics is not the finding.

The political record is where the loudest claims about both regimes live, and where the evidence is weakest for them. It is assessed at four separate levels, because an outcome that falls unevenly is not evidence of intent.

Level
Twitter
X
LEVEL 1Workforce and ownership
Twitter

A left-leaning workforce culture, acknowledged by the chief executive in 2018. Employee-affiliated political donations were about 99% Democratic in the 2020 cycle (FEC-derived figures via secondary reports).

X

A controlling owner who became a declared partisan actor: endorsed Trump (13 Jul 2024), supplied about 91% of America PAC's declared contributions, advised the President (Jan–May 2025), then broke with him (Jun 2025). An owner with direct power over ranking and enforcement is different in kind from a diffuse donation skew.

LEVEL 2Asymmetric outcomes
Twitter

Users of pro-Trump hashtags were 4.4 times as likely to be suspended in 2020–21, and also shared lower-quality links, even as rated by Republican-only panels (Nature, 2024). Behaviour explains the gap: Medium-High.

Twitter's own randomised experiment found the mainstream right received more algorithmic amplification than the mainstream left in six of seven countries (PNAS, 2021).

X

Right-leaning reach rose in measured episodes: contentious far-right accounts after the takeover (2023); a break in the owner's and Republican-aligned accounts' reach from 13 Jul 2024 (pre-print; cause unestablished); a right-leaning tilt in new accounts' default feeds (2025).

Whether this is algorithmic or audience-driven: Medium, since Twitter's ranking already favoured the mainstream right.

LEVEL 3Individual decisions
Twitter

The laptop-story block (reversed within about 48 hours), the Trump suspension (with a published rationale), the Babylon Bee lock. Takedown requests reached Twitter from both the Biden campaign and the Trump White House.

X

Journalists suspended under a same-day rule (Dec 2022); a left-of-centre journalist suspended over a hacked campaign dossier (26 Sep 2024), formally mirroring the 2020 laptop case; reinstatements by poll with no stated standard. Enforcement also reached right-wing accounts.

LEVEL 4Systematic partisan intent
TwitterNot established

No internal directive to enforce by viewpoint surfaced, including in the Twitter Files. Medium-High · bounded negative

XNot established

The evidence shows episodic, personal owner intervention, not an institution-wide programme. Low-Medium that a programme existed

What the political record adds. It is not a scorecard. The claim that Twitter systematically censored conservatives is not established; neither is the claim that X systematically manipulates the platform for them. What the record does show is the shift found everywhere else on this page: from institutional edge-case error under Twitter to personal owner intervention under X.

10

Mechanism ledger

Seven mechanisms. Six were already there.

Each verdict weighs severity, affected population, dependence, notice and remedy, not the number of incidents.

Mechanism
Where
Verdict
Confidence
Remedy Asymmetry
Present under both

Materially intensified under X in contractual terms. EU statutory remedies materially complicate a universal adverse finding.

High on termsIn practice: unresolved
Contractual Discretion
Present under both

Core legal discretion substantially continuous; changed mainly in beneficiary and locus. Materially intensified for paid, creator and developer relationships.

High
Ecosystem Extraction
Present under both

Real Twitter antecedents, 2011–2018. Materially intensified under X in severity, population, speed and enforcement.

High
Hidden Degradation
Present under both

Changed mainly in target; net adverse under X. X materially improved disclosure of rule-based visibility limits, and Twitter's 2FA-data misuse is the largest single hidden harm in the record.

Medium
Forced Monetization
Present under both

Materially intensified under X. Five components new in kind: sale of status, SMS 2FA behind payment, paid eligibility to earn, read limits by payment, the $15,000 tariff.

High
Pay-for-Visibility
Present under both

Intensified 2022–24; beneficiary moved from editorially selected to paying accounts. Paid reply prioritisation persists; a paid For You boost in 2025–26 is unresolved.

High 2022–24For You 2025–26: Low-Medium
Customer-to-Adversary Conversion
Present primarily under X

Advertisers. Critics and researchers: facts established but adjacent. Competitors and publishers: Medium. Subscribers, creators, developers: unresolved or absent.

High for advertisers
11

Final synthesis

What actually changed.

Twitter supplied much of the legal and technical substrate. X changed seven things about how it was used.

WhoThe owner, in public, in place of a senior group working through written policy.
How fastSame-day rules, zero-notice cut-offs, 30-day repricing.
For whomPaying tiers, the owner, affiliated AI products, litigation strategy.
What was pricedStatus, a security control, reach, the right to earn, reading, data access.
EcosystemsDevelopers, researchers, publishers and creators squeezed in months, not a decade.
RemediesFrom California courts or opt-out arbitration to Texas courts and class waivers, outside the EU.
ResistanceCustomers who stopped buying were sued.

X did not need to invent a new system of platform power. Twitter had already built much of it. What changed was the restraint around that power: who exercised it, how quickly, who benefited, what customers had to pay for, and what happened when some of them simply walked away.

Twitter was not innocent. It wrote the discretion clause, built undisclosed reach reduction, squeezed its developers for a decade and paid the largest user-data penalty in the record. X was not merely Twitter with a different owner. It put status on sale, moved a security control and the right to earn behind payment, rewrote user remedies four times in 31 months, and sued customers who stopped buying.

The reading the evidence supports is layered: continuous in the substrate of discretion; discontinuous in the monetisation model, the remedy architecture, the treatment of resisting counterparties and the locus of discretion, with one mechanism, Customer-to-Adversary Conversion toward advertisers, for which no Twitter antecedent of comparable kind was located. Some transparency improved even as discretion intensified elsewhere, and that is part of the finding, not a footnote to it.

Method & limits

How this was established.

Evidence base and cutoff

The evidentiary cutoff is 2 September 2026, 23:59 UTC. The investigation synthesised ten independent clean-room research streams (moderation, verification, subscriptions, visibility, developers, advertisers, creators, data access, terms and remedies, political asymmetry) under a common brief that forbade reliance on earlier X/Twitter research. The primary base includes 27 archived captures of the Terms of Service (Jun 2020 – Apr 2026) plus a 2016 copy; the Purchaser, Creator Revenue Sharing and Subscriptions Creator Terms; the git histories of twitter/the-algorithm and xai-org/x-algorithm (to commit 85ac72a, 2 Sep 2026); European Commission DSA documents; X's transparency reports; court rulings as reported; and peer-reviewed studies. A 172-claim audit and three adversarial reviews preceded publication.

Several archives and news outlets were unreachable during research, so some pre-2022 help-centre wording could not be verified. Where a claim rests on secondary reporting, this page says “reported” and carries the original confidence rating.

Events after the cutoff are not evidence

Research encountered later events and excluded them. They include the retirement of X's Creator Revenue Sharing programme on 7 Sep 2026 and the launch of its replacement; a Terms of Service version effective 9 Oct 2026 (first captured 10 Sep 2026); ranking-code commits after 2 Sep 2026; and an 8 Sep 2026 ruling in a separate case. None is used on this page.

Why there is no CHI score

This is a comparative investigation of two regimes of one institution, not a company assessment. No locked score exists for X, and CHI does not derive one from mechanism counts. The page reports verdicts and confidence levels instead.

What this page does not claim
  • That Blue subscribers received four times or twice the reach. The 4.0 and 2.0 values are code defaults.
  • That the announced verified-only For You rule shipped, or that paid verification changed moderation outcomes. Both are unresolved.
  • That X issued no refunds. No company-issued refund was located in the reviewed record.
  • That the advertiser suit was thrown out in full. Dispositions differed by defendant, and the WFA claim survived to settlement.
  • That a court ordered X to stop training Grok on EU data. X agreed to suspend processing in Irish High Court proceedings brought by the Data Protection Commission (Aug 2024).
  • That either regime ran a systematic partisan enforcement programme, or that the owner's politics caused the July 2024 ranking shift.
  • That X invented these mechanisms, or that Twitter was benign.
What remains unresolved
  • Live values, and the removal date, of the 2023 Blue-author multipliers.
  • Whether a paid For You boost operated in production in 2025–26.
  • Whether paid status affected moderation outcomes.
  • Refunds to November 2022 checkmark buyers, ad-free-articles subscribers and third-party-client subscribers.
  • Whether creator payouts rose or fell after November 2024.
  • The Ninth Circuit's disposition of X v. CCDH, and the Fifth Circuit appeal in the advertiser case.
  • Whether the July 2024 reach shift was algorithmic or audience-driven.
  • Which regime gave users the more effective remedy in practice.

Continue

Related CHI work

Special InvestigationElon Musk Asymmetric Commitment across Tesla, X, Starlink and xAI: what is said, what is signed, and who bears the gap. IndexSpecial Investigations CHI's deeper research projects, beyond a standard company assessment. How CHI worksMethodology Evidence grading, confidence levels and the disciplines this page applies. LexiconHostility patterns The CHI vocabulary for the customer-facing patterns this investigation touches.