Images
@jab1 wrote:
My local (Sheffield) forum is considerably smaller than this one ... has had automated image verification for at least a year.
If one man, with the help of the platform vendors, can sort the issue, why is it so hard for the BT Group PLC behemoth?
Grok suggests -
Understanding the Challenge and the Online Safety Act
The UK's Online Safety Act 2023 (OSA) imposes duties on online platforms, including user-to-user services like forums, to proactively identify and mitigate risks from illegal or harmful content. This includes images that may contain child sexual abuse material (CSAM), terrorist content, violence, or other prohibited material. For platforms hosting user-generated content, such as Plusnet's forum, the Act requires risk assessments and effective moderation systems, with codes of practice for illegal harms effective from March 2025. Manual approval processes, while compliant, are inefficient and can hinder user engagement, as you've described. AI image recognition offers a scalable, near real-time alternative that aligns with OSA requirements by automating detection while allowing human oversight for edge cases.
Proposed AI-Powered Solution for Automated Image Approval
Plusnet can integrate AI-based content moderation into their forum's upload workflow to scan images automatically upon posting. This would enable most safe images to be approved and displayed in seconds, reducing wait times to near real-time (typically under 1-2 seconds per image via cloud APIs). The system would flag potentially harmful content for human review, ensuring compliance without bottlenecking legitimate posts.
How the AI Image Recognition Process Would Work
- Image Upload and Initial Scan: When a user posts an image, the forum software (e.g., via plugins for platforms like Discourse or phpBB, which Plusnet might use) intercepts the upload and sends the image to an AI moderation API. The AI uses computer vision models trained on vast datasets to analyze the image for:
- Illegal/Harmful Content: CSAM (using perceptual hashing or classifiers like those from Thorn or PhotoDNA), explicit nudity, violence/gore, weapons, hate symbols, or terrorism-related imagery.
- Contextual Risks: Even non-illegal content harmful under OSA, such as self-harm promotion or bullying visuals, if the AI model supports it.
- Detection is probabilistic: Models output confidence scores (e.g., 0-1 scale) for categories like "adult content" or "violence."
- Automated Decisioning:
- Approve Instantly: If the AI confidence score indicates safety (e.g., >95% safe across all categories), the image is released and displayed immediately. This covers the vast majority of benign posts (e.g., screenshots of router setups or memes).
- Flag for Review: If flagged (e.g., moderate confidence in a risky category), the image is held in a queue for a moderator. Notifications can alert staff, with tools for quick bulk approvals/rejections.
- Reject Outright: For high-confidence violations (e.g., clear CSAM), auto-reject and notify the user/admin, logging for OSA reporting.
- Near Real-Time Performance: Cloud-based AI services process images asynchronously but return results in milliseconds to seconds. Integration via APIs ensures the forum doesn't freeze during scans. For high traffic, use queuing systems like AWS Lambda or Azure Functions to handle bursts.
- Fallback and Logging: All scans are logged (anonymized where possible) for audits, as required by Ofcom (the OSA regulator). If AI fails (e.g., due to downtime), revert to manual hold.
This hybrid approach (AI + human) meets OSA's "systemic risk mitigation" duties while minimizing delays.
Recommended AI Tools and Technologies
Several mature, OSA-compliant AI services are available, often with built-in support for UK regulations. Plusnet should select based on ease of integration, cost (typically pay-per-scan, e.g., $0.001-0.005 per image), and accuracy (aim for >99% for safe content). Here's a comparison:
Start with Amazon Rekognition or Azure, as they offer robust CSAM detection and are widely used by UK platforms. For open-source alternatives (if cost is an issue), consider integrating models like OpenAI's CLIP or Hugging Face's safety classifiers, but these require more dev work and may not be as compliant out-of-the-box.
Implementation Steps for Plusnet
- Assess and Plan (1-2 months): Conduct an OSA risk assessment (required by law) focusing on image harms. Engage a compliance expert (e.g., via Ofcom guidance). Choose a tool and test for false positives on sample forum images.
- Technical Integration (2-4 weeks):
- Modify forum upload handler (e.g., via hooks in the CMS) to call the AI API on post-submit.
- Use secure endpoints (HTTPS) and store images temporarily in a blob store (e.g., AWS S3) during scanning.
- Implement user feedback: Show "Image under review" placeholders that auto-update.
- Testing and Rollout (1 month): Beta test with a subset of users. Monitor metrics like approval rate (target 90%+ auto-approved), false positive rate (<1%), and scan time. Train moderators on AI outputs.
- Ongoing Maintenance: Regularly update AI models (providers handle this). Report incidents to Ofcom as needed. Comply with ICO privacy guidance by minimizing data retention (e.g., delete scans after 30 days).
Total setup: 2-3 months, with minimal ongoing staff needs.
Benefits
- User Experience: Near-instant approvals restore forum vitality, encouraging help threads without delays.
- Compliance and Efficiency: Reduces moderator workload by 80-90% for images; ensures proactive OSA adherence, avoiding fines (up to 10% of global revenue).
- Scalability: Handles growth without hiring; AI improves over time with provider updates.
- Cost-Effective: For a forum like Plusnet's (likely <10,000 images/month), annual costs could be under £5,000.
Key Considerations and Challenges
- Accuracy Trade-offs: AI isn't perfect—false positives (e.g., flagging medical images as explicit) could frustrate users. Mitigate with tunable thresholds and appeals processes. False negatives are riskier for compliance, so err toward caution.
- Privacy and Ethics: Scanning images processes personal data, so obtain consent via terms of service and use UK/EU data centers for GDPR. Avoid storing full images post-scan.
- Bias and Coverage: Models may underperform on diverse content (e.g., non-Western imagery); test thoroughly. For OSA-specific harms like disinformation visuals, combine with text moderation.
- Legal/Regulatory: Consult lawyers for OSA implementation statements (due to Ofcom). If using third-party AI, ensure SLAs cover uptime and liability.
- Edge Cases: AI struggles with altered images (e.g., deepfakes); supplement with hashing for known CSAM databases.
This AI approach transforms a "hopeless" manual process into an efficient, compliant system. Plusnet could pilot it quickly to demonstrate commitment to user safety and engagement.
So having been told that it is 'unreasonable' to expect prompt image approval after 10pm
and that the best way to speed up an image approval is to report my own post to the moderators for action.
I posted Reply #81 with an embedded image, at 3:42pm, and immediately reported the post to the moderators.
Two hours and ten minutes later, the image gets approved.
HOWEVER in the mean time 79 people attempted to view my post, and were unable to view the most important part of my posting, and ARE UNLIKELY to ever go back and re-read the long post or notice that the all important table has now become visible.
This is useless, the image approval needs to be less than five minutes, otherwise all those people subscribed to a topic will eventually give up responding to their topic update notifications, if parts of posts are frequently missing.
