AI & Email Immutability: A Hacker News Discussion

Our staff is encouraged to understand and safely utilize existing tools. Company policy dictates that all tools, including those using AI, must adhere to privacy-preserving principles.

Data Protection

All data protection, confidentiality, and privacy policies must be followed. Vendors for anti-abuse and support are adopting AI for translation, categorization, and abuse detection, and we ensure their policies protect customers.

Accountability

Any AI-generated writing or code requires review and understanding by a human, and it also must go through secondary review before being used.

Bias Awareness

Actively seek out biases or hallucinations in AI outputs.

Human Authority

A path to appeal decisions made by automated tools must be available.

While plain text remains unchangeable, contemporary emails often incorporate remotely hosted elements. Upon opening such emails, the relevant content is fetched directly from the sender. This remote content isn't permanently stored but rather cached temporarily and isn't guaranteed to be reused if the email is accessed much later. If these elements vanish or are altered, email providers have limited recourse. Dynamic AMP messages can also change over time.

The sender can't treat it as a read receipt, because the provider can state "Our infra performs this operation for the user for immutability purposes" similar to other email operations that proxy these requests for privacy purposes. Seeing an email update with tracking progress can be interesting; however, receiving a new email is often preferable.

Although probably no one outside the company ever read the plaintext version of the email, it proved invaluable for verifying content accuracy by outputting the complete message contents to the console. While a plain text email may not be widely appreciated, particularly outside a niche audience, the underlying email formatting remains relatively unchanged over the past five decades. However, complexities arise when attempting elaborate designs.

HTML email is vulnerable because linked content could be hosted anywhere. Text-only email remains king and is still used in many contexts. Apple's private loading feature offers a potential solution by having the mail server retrieve and save referenced content to ensure consistent rendering of the message.

One's choice of email provider is not necessarily based on concern about the aforementioned potential changes. Cameras documenting news will likely require watermarking, fingerprinting, and authentication, similar to current practices by Canon, Nikon, and AFP. Trustworthy visuals may soon be limited to organizations employing such measures. Social media platforms might promote direct camera-to-post functionality for verifying unaltered content. For example, videos of protests filmed through the Instagram app could receive a "this is real" badge, whereas uploaded videos would be flagged as potentially AI-generated.

While this may not impact the original source, AI may become adept at detecting screen recordings. Manipulating media, including documents, has become easier. Courts have historically addressed manipulated evidence. Lawyers must carefully vet jurors who could be swayed by AI manipulation. Watermarking and platform-based clickable watermarks are important. The general public doesn't always prioritize mainstream media, which poses a bigger threat than AI videos. Concerns about app sideloading are minor in comparison to the impact of manipulated media on conflicts.

Common online posts consist of images with text, tweets, or headlines, which are often taken at face value without verification. Requiring camera-signed photos may not significantly curb the spread of misinformation if a simple image of text is enough for people to believe it. This concentrates power with the platform.

Audiences will need to learn to trust information in new ways; technology alone isn't a complete solution. LIDAR sensors could offer another way to verify image depth information.

It seems unlikely Instagram, owned by someone keen on AI content, would embrace such verification methods.

Cameras may record screens showing AI-generated videos and authenticate them. While screen pixel mapping might reveal such recordings today, dedicated screens will emerge. Adding LIDAR is suggested as a countermeasure, but mirroring the sensor could bypass it. At the end of the day, personal responsibility is key. People must assess claims critically.

It's already possible, citing The Mandalorian's Stagecraft studio as an example. Stagecraft uses LED walls to display dynamic digital sets, allowing real-time reactions and content manipulation during production. These high-definition LED video walls display computer-generated imagery, previously added during post-production.

Proving the integrity of an email to others presents a greater challenge. One would need to retain DKIM keys from the signing server to validate historical DKIM signatures. It would be important to verify the old message was not forged or altered. DKIM signing keys may be logged in a Certificate Transparency log, enabling verification of a domain's DKIM key history. There are some tools but they are incomplete.

Email historically lacked built-in security. The article notes, "An email is your copy, and the sender can’t revise it later." However, you can only be absolutely certain no one is tampering with your email if you don't share it with anyone else. One user expressed a desire for Fastmail to avoid adding AI features.

Domain age, IP address, and DKIM/DMARC compliance matter. Switching to hosted email resolved delivery problems. Sending an initial email and then replying prevents being blocked. Use mail-tester.com to confirm proper DKIM setup and lack of IP blacklisting before emailing. One person uses everything they can to block trackers, spy ware, etc and have never been "Cloudflare blocked".

Google's AI overview consistently outperforms a proprietary AI assistant.

While Fastmail allows adding users to billing plans, each user requires a separate paid plan, which is fine for heavy users but not ideal for family members. There is no multi-user pricing.

After paying for Gemini, storage usage decreased. Consider switching to Fastmail. Bayesian spam filtering categorizes 99% of email as noise. Current AI usage is limited by employers opposing privacy enhancement and ad filtering.