Welcome back! Legal work, big tech budgets, lab science and music creation all run into AI at once. One tool asks law firms and finance teams to trust it with real documents, another leans on huge cloud and chip spending, a research system suggests new materials from decades of chemistry papers, and a music generator turns ideas into instant soundtracks. The tension is simple: how much paid, expert work we are willing to move from people to software.
In today’s Generative AI Newsletter:
Anthropic legal AI tool hits software stocks.
Microsoft spends $37.5B but Copilot struggles.
MIT uses AI to discover new materials.
AirMusic turns simple text prompts into songs.
Latest Developments
Lawyers Face AI Pressure as Anthropic Causes a $300B IT Stock Drop
Anthropic’s new Claude Cowork Legal automation tool spooked investors as soon as it hit the market, crushing software and data stocks. The fall began when reports linked the release to agent style plugins inside Claude that contract review, regulatory workflows and legal briefings, which are typically associated with costly subscriptions, framing it as AI that can do real office work. The market reacted strongly, with estimates indicating a loss between $285B and $300B. Concerns also affected India's IT services stocks, as these companies rely on billable staffing for revenue.
Here’s what the evidence points to:
Scope: The tool targets legal workflows, but Anthropic said that it does not provide legal advice and requires attorney oversight.
Damage: Several major players fell significantly, including Gartner, Thomson Reuters, Relx and Wolters Kluwer.
Spillover: The slide hit private fund managers like Ares, KKR and Blue Owl because they own large chunks of software businesses.
Tradeoff: Buyers get faster review and lower costs, but errors create a real risk of new lawsuits.
The upcoming challenges will focus on liability, data rights and whether established companies integrate AI quickly enough to prevent their profits from diminishing. AI is learning to move work through systems, making seat-based software feel overpriced overnight. While legal departments may save time and money, the industry must still address questions about liability, data accessibility, and maintaining trust when the tool makes a confident error. This tension is expected to influence the development of future AI products and subsequent stock sell-offs.
Microsoft’s Copilot Struggles: Only 3.3% of Users Pay for AI
Microsoft developed Copilot to serve as the AI feature in Office, Windows, Teams, and other Microsoft products that users already pay for. In recent months, many users experimented with Copilot through the default AI plan and then reverted to their previous preferences. Microsoft invested $37.5B in AI-related infrastructure in the most recent quarter. However, only 15M out of 450M Microsoft users purchased Copilot spots. Additionally, Copilot's primary tool usage decreased from 18.8% in July to 11.5% in January.
Here’s what the details suggest:
Conversion: Computing says only “3.3%” of Copilot Chat users converted to paid.
Price: Microsoft set the add-on at $30 per user per month, a clean number that feels less clean in procurement meetings.
Pressure: The internal pressures, such as training and performance nudges, indicate that the product requires assistance in marketing itself.
Defense: Microsoft's security team is enhancing security measures for an 'AI-powered world,' highlighting risks such as prompt injection and data poisoning.
Copilot has a competitive edge because Microsoft dominates most work environments. We have seen this pattern before, where big companies bundle a new feature while startups establish user habits. If it can make the product feel essential by tangible time-saving benefits in Word, Excel, PowerPoint and Teams, the price will become easier to defend and with better renewals. If it keeps feeling unreliable or expensive, it risks being perceived as the costly but non-essential feature in the budget.
MIT Trains AI to Develop New Complex Chemical Materials
Researchers at MIT are working to help labs make materials quickly and effectively instead of keeping them virtual. Their tool, DiffSyn, examines 23,000 lab recipes across 50 years to provide detailed instructions on crafting complex materials such as zeolites (tiny sponge-like rocks, made of silicon and oxygen and used as filters and fuel-making). The lead author, Elton Pan, describes the current reality as 'domain expertise and trial and error,' meaning that experienced individuals continue to invest significant time in trial and error processes.
Here are the key claims:
Speed: Pan says scientists test recipes individually, but DiffSyn can evaluate 1,000 in less than a minute, allowing rapid exploration.
Method: MIT defines this process as a diffusion approach that converts noise into meaningful structure.
Proof: The study shows that a new zeolite called UFI was successfully made by following the steps suggested by the model for better heat tolerance.
Risk: The model relies on published papers, which may omit unsuccessful attempts, leading to the continuation of flawed practices.
This points to where the AI industry is heading next. Models are moving from text and image generation into operational decisions in everyday situations. It is dangerous when the model can output 1,000 options faster than humans can assess them. Those with accurate experimental records and fast feedback loops for testing suggestions will outperform zeolites. DiffSyn looks useful, but it also nudges labs toward high-speed experimentation, which can burn time and materials just as quickly as it saves them.
AirMusic: Make Simple Background Music in Minutes
AirMusic is a website that helps you make music by typing what you want. You can create a short track, make it longer, or pull a song apart into pieces like vocals and drums. It is useful when you need clean background music for a video, podcast, ad, or demo and you want to finish fast
Core functions (and how to use them):
Text to music: Type what you need, like “60-second chill beat, no vocals," and download the audio for your edit.
Extend a track: Take a short clip and extend it so it fits your exact video length, like 30 seconds or 60 seconds.
Stem separation: Split a song into multiple parts so you can mute vocals, keep only drums or pull out the bass for a remix.
AI voice cloning: Upload vocals and generate a quick cover in another voice to test an idea before recording.
Lyrics to vocals: Paste lyrics and generate a basic vocal take so you can hear how the hook and chorus feel.
Try this yourself:
Make a 60-second background track for a piece of content. In Create, type “upbeat pop instrumental for a product demo, 120 BPM, no vocals.” Download the track and drop it under your video. If your video is not exactly 60 seconds, use the extend tool to match the length. Then run the same track through the split tool and turn down or remove any part that clashes with your voiceover.
