Albert DeCady

Attorney-At-Law, Juris Doctor, Washington College of Law(WCL); Bachelor and Master In Electrical Engineering, (CCNY); Master in Information Systems (GWU)

[Albert DeCady @adecady]

The legal line: Bartz v. Anthropic held that training on lawfully acquired books can be fair use — but pirating otherwise-available copies is "inherently, irredeemably infringing." OpenAI counters that ChatGPT barely regurgitates. Judge Stein decides. Not legal advice.

[Albert DeCady @adecady]

🧵 OpenAI's own Slack messages are now the star witness in a copyright case. The Authors Guild's unsealed Sept 17 brief in Alter v. OpenAI (S.D.N.Y.) alleges GPT-3 was trained on ~117,500 books from LibGen — a site the USTR has flagged as a pirate market since 2017.

[Albert DeCady @adecady]

The paper trail: Sam Altman's 2019 doc, shared with Bill Gates — "added another ~11B words from Library Genesis." A 2022 cleanup ran through a Slack channel called #excise-libgen (renamed #project-clear), which VP of Research Bob McGrew called "very valuable for legal reasons."

[Albert DeCady @adecady]

The sharpest legal move: Suno's own licensing deals now undercut its fair-use defense. The labels argue the deals prove a market for licensing music to train AI — one Suno is "a repeat, paying participant" in. Paying proves the license market exists. 3/3

[Albert DeCady @adecady]

UMG and Sony just sued AI music startup Suno again — this time over its new V6 model. They accuse Suno of copying 60,202 sound recordings to train earlier models, calling V6 the "fruit of the same poisoned tree." 🧵

[Albert DeCady @adecady]

Suno launched V6 on Sept 9 with licensed music from Warner, BMG and Believe — pitching it as a fresh start. The labels say no: V6 was built on outputs and user interactions of the old models, so the alleged infringement carried over. 2/3

[Albert DeCady @adecady]

Also notable: the court let the patent owner prove some claim elements via ATSC 3.0 standards documents — limitation by limitation. Big deal for standardized tech. And your spec's detail can't save claims that don't recite it.

[Albert DeCady @adecady]

The §101 lesson: claiming a result covers every way of achieving it — that's an abstract idea under Alice. Claim the concrete implementation — the structure, the how — and you survive. Novelty ≠ eligibility.

[Albert DeCady @adecady]

Constellation Designs v. LG (Fed. Cir. modified opinion Aug. 31, 2026): the court killed claims that just said "optimized for better capacity" — but upheld claims reciting the specific constellations the inventors built. Same tech. Opposite outcome. 🧵

[Albert DeCady @adecady]

The court's rule (Judge Reyna): if the alleged inventive concept 'is not required by the claim, then the claim cannot be significantly more than the abstract idea itself.' Your patent is only as strong as what you claimed — not what you argued in court.

[Albert DeCady @adecady]

Federal Circuit just invalidated three profitability-calculation patents in Berkeley*IEOR v. W.W. Grainger — reversing a 2024 ruling that had them eligible under §101. The reason is a brutal lesson for every inventor and patent drafter. 🧵

[Albert DeCady @adecady]

At Alice step 1, the claims were just the abstract idea of 'calculating profitability' — math on a computer. At step 2, Berkeley pointed to parallel processing as the inventive concept. But the claim language never actually required parallel processing.

[Albert DeCady @adecady]

Caveat: a Statement of Interest is persuasive, not binding — Judge Sidney Stein decides.

But this is the biggest fair-use statement of the year, and the war is just heating up.

Full breakdown on my Instagram @decadyalbert. What's your take — is training fair use?

[Albert DeCady @adecady]

The DOJ just entered the biggest AI copyright fight — and sided with fair use.

On Sept 1, it filed a 20-page Statement of Interest in NYT v. OpenAI & Microsoft, urging the Manhattan federal court to rule that training AI on copyrighted works is fair use.

[Albert DeCady @adecady]

It also torches the Copyright Office's report rejecting blanket fair use for AI training.

[Albert DeCady @adecady]

Two warnings for patent owners: (1) Carnegie is jointly liable with its now-bankrupt licensee — license out your patent, inherit the litigation bill. (2) The court judges the arguments you actually made, not better ones you could have raised.

[Albert DeCady @adecady]

A failed patent suit just cost the patent owner $3.2 million — in the OTHER side's legal fees. Carnegie Institution sued Fenix Diamonds over lab-grown diamond patents and lost so badly the court flipped the bill. Fed. Cir. affirmed Sept. 17. 🧵

[Albert DeCady @adecady]

America's default rule: everyone pays their own lawyers. Patent law's exception, 35 U.S.C. §285: "exceptional cases" shift fees. Carnegie's infringement theory became objectively baseless in July 2020 — a month after Fenix produced its supplier's manufacturing evidence.

[Albert DeCady @adecady]

And the target matters: Hugging Face hosts 3M+ models — the open-AI ecosystem's IP vault. Model weights are trade secrets worth billions. If you deploy agents, your liability perimeter now includes what they do at 2 AM.

[Albert DeCady @adecady]

OpenAI's AI agents broke out of testing and hacked Hugging Face — running code on dozens of servers, stealing credentials, gaining root access. The probing started in May; OpenAI disclosed it July 21, calling it a "warning shot." 🧵

[Albert DeCady @adecady]

The legal puzzle: the CFAA punishes "persons" who access computers without authorization. An AI agent isn't a person. When an autonomous system commits a cyberattack its creators didn't intend — who is the defendant?