Photosensitive Epilepsy Risk
High-frequency animation and strobing content can trigger seizures in susceptible users. Current upload systems have no preemptive protection mechanism.
Executive summary
Ship the creative feature. Build the safety layer first.
If X introduces dynamic profile headers and sound reactions, the creative upside is obvious. So are the new risks in accessibility, abuse prevention and regulatory exposure that static images do not carry. This proposal is written for the product and engineering teams who would own that launch; X has announced no such feature, and nothing here claims otherwise.
This is not about restricting creative features. It is about building the minimum safety and compliance infrastructure the platform will eventually need — before it becomes expensive or harmful to users.
This framework establishes a practical governance layer to allow dynamic headers to ship responsibly. It prioritizes three core objectives:
- Protecting users from photosensitive epilepsy risk through mandatory pre-upload analysis
- Reducing abuse surface with technical guardrails and identity-linked accountability
- Meeting WCAG 2.2 accessibility guidance and obligations under the EU DSA and EU AI Act before enforcement creates liability
What you get, if you are the team that would build this
- Upload-time controls with concrete numbers: 15 fps, 10 s loops, 8 MB, 3 flashes a second (general and red), 8 s reactions, 140 s voice posts, and an eight-check preflight. All of them are editable policy, not hard-coded opinion.
- A working reference implementation in the browser: the flash scan (WCAG 2.2 general and red-flash rules, CIE 1976 UCS), the header player rules, the reaction and voice-post players, a review queue and an audit log, with unit and end-to-end tests you can run from the repository.
- An honest boundary. Abuse and copyright checks are labelled SIMULATED, the scan is a screen rather than a certified analyser, and the identity layer is a proposal. The places that need X's own services are marked, not papered over.
01 · The real problem
Motion changes the risk profile.
Static profile headers carry relatively low operational and regulatory risk. Dynamic and interactive media fundamentally change the risk profile in three material ways:
Abuse Vector Expansion
High-velocity, unverified media uploads create new attack surfaces for harassment, impersonation, and coordinated abuse campaigns.
Regulatory Exposure
EU DSA obligations already apply to Very Large Online Platforms like X, and the EU AI Act adds transparency rules for AI-generated audio and video (Article 50).
02 · Proposed framework
Three layers, each one load-bearing.
A layered governance approach designed to balance creator velocity with user safety and regulatory compliance.
Layer 1
Media Safety Screening
Mandatory Photosensitive Epilepsy Analysis (PEA) before any dynamic media is published.
Layer 2
Technical Guardrails
Hard constraints on file size, frame rate, duration, and playback behavior.
Layer 3
Identity & Accountability
Strictly opt-in biometric anchoring with a clear, proportionate enforcement model.
03 · Core controls
What actually runs at upload time.
03.1
Media Vetting & Safety Screening
All dynamic media must pass a Photosensitive Epilepsy Analysis (PEA) before publication. This is a hard blocking control.
- STEP 01PEA Scan10–20 seconds
- STEP 02ResultPASS / FAIL
- STEP 03Provisional Publish+ Warning Label
- STEP 04Manual Review24hr SLA
- Hard BlockSevere cases only
Scan Time
10–20s
Publish Gate
PASS required
Flash Threshold
≤ 3 per second
Red Flash Rule
R/(R+G+B) ≥ 0.8
The scan is a simplified screen based on WCAG 2.3.1’s general and red flash thresholds, not a certified photosensitive epilepsy analysis. A PASS means the clip stayed under these limits in this check; it is not a safety guarantee.
03.2
Technical Guardrails
Maximum File Size
8 MB
Maximum Frame Rate
15 fps
Maximum Loop Duration
10 seconds
Playback Behavior
Auto-pause on scroll + low power mode
All uploads are transcoded server-side from GIF to WebM/VP9. We expect this to cut file size by roughly 60–80% (an estimate from typical GIF-to-VP9 results, not a measured benchmark) and to improve playback on lower-end Android devices.
03.3
Identity & Abuse Prevention
Dynamic media carries meaningfully higher abuse risk than static images. This framework uses a strictly opt-in identity anchor through an ID-verification vendor such as AU10TIX; whether X already uses one for this purpose is not assumed here. Users who decline verification retain full access to static headers and all non-dynamic profile features.
Three-strike enforcement model
-
Strike 1
Warning + temporary restriction on dynamic uploads
-
Strike 2
12-hour suspension of dynamic media upload capability
-
Strike 3
Permanent suspension reserved exclusively for verified malicious patterns (automated probing, coordinated attacks). Standard users receive a 7-day suspension + mandatory compliance review.
04 · Staged rollout
Expand only when the gate holds.
-
Premium+ Subscribers
Limited 30-day rollout. Monitor CPU/GPU load, false positive rate, and user complaints.
30-day gate -
All Verified Accounts
Expanded rollout with strict A/B testing. Gate: complaint volume below 0.5%.
<0.5% complaints -
General Availability
Full platform rollout with automatic low-power mode and reduced bitrate for limited-data users.
Full availability
05 · Risk & limitations
What this framework does not solve.
This framework is intentionally scoped. The following limitations are acknowledged:
Client-side rendering attacks
This framework governs upload-time controls. Malicious client-side injection after scanning requires platform-level sandboxing.
Steganography
Automated PEA scanning has known technical limitations against hidden triggers. Behavioral signals provide a secondary detection layer.
Biometric data concentration
Scoped exclusively to dynamic media uploads. Static headers remain unaffected. Existing 30-day retention policies apply.
Artistic false positives
The provisional publish model was specifically designed to protect legitimate creators from overly aggressive automated blocking.
06 · Regulatory alignment
Mapped precisely, claimed carefully.
WCAG 2.2
Used as the accessibility benchmark, including the general and red flash thresholds. ADA Title II covers state and local governments, so it is not cited as binding on X.
BenchmarkEU Digital Services Act
Articles 16, 17 and 20 (notice and action, statements of reasons, internal complaints) mapped to header moderation; Articles 34 and 35 to systemic risk assessment and mitigation.
Art. 16 · 17 · 20 · 34 · 35EU AI Act
Article 50 transparency rules applied to AI-generated or AI-altered media. The Article 14 human-oversight rules cover high-risk systems, not content moderation, so human review here is a design choice, not a legal requirement.
Art. 50NIST Privacy Framework 1.0
Govern, Control, and Communicate functions mapped to dynamic media workflows.
Govern · Control · Communicate07 · See it working
Dynamic headers and sound reactions, governed by this framework.
A browser-only concept prototype of the two features this framework covers. A dynamic GIF profile header that must pass the photosensitivity scan and guardrails before it goes live, and short sound reactions that never autoplay, carry captions and pass a preflight check.
Sound Reactions · concept prototype
See a real GIF banner pass the scan
The banner on the right is a real GIF playing straight from GIPHY, the way X serves GIFs. The prototype fetches it and runs the flash scan in your browser before it is allowed to play; an earlier pass on the owner’s copy found 14 fps on a 1.3-second loop, 0.5 MB and 1 general flash per second, under the limit of 3. Two more real GIFs play from Tenor on other profiles; one runs at 20 fps and is drawn at the 15 fps cap. Open a profile, tap Edit profile and scan the other samples or a GIF or video of your own (it never leaves your browser) to see which are blocked, rejected or published with a warning. Tap any name in the feed to see that person’s banner, then try sound reactions. All other accounts are fictional; the abuse and copyright checks are labeled as simulated.
Open Sound Reactions
Real GIF · via GIPHY · PEA passed
Not affiliated with X. Source code is in the sound-reactions-src folder of this repository.