Cutting Through AI Coding Hype (Without Cutting Yourself) — and the Biotech Trends Worth Watching in 2026

AI generated image for Cutting Through AI Coding Hype (Without Cutting Yourself) — and the Biotech Trends Worth Watching in 2026

On January 16, 2026, MIT Technology Review published an edition of its weekday newsletter The Download titled “The Download: cut through AI coding hype, and biotech trends to watch”, credited to Rhiannon Williams. The piece is a compact two-for-one: a reality check on AI coding’s claims of effortless productivity, and a quick scan of biotech currents worth paying attention to as the year gets rolling.

Consider this article a long-form, caffeinated expansion of those themes—written for the people who have to ship software and fund science, not just tweet about it. We’ll separate hard evidence from vibe-based marketing, map the risks that show up once AI starts writing production code, and connect the dots between “AI for coding” and “AI for biology”—because in 2026, those are increasingly the same conversation.

Original RSS source: MIT Technology Review, The Download (January 16, 2026). Original author/creator: Rhiannon Williams. Read the original here.

AI coding: why the hype is so loud (and why your backlog is still there)

AI coding tools are the rare enterprise product that can be demoed in under 30 seconds and sold in under 30 minutes. Type a prompt, get a function. Paste a stack trace, get a fix suggestion. Ask for tests, receive tests. The dopamine hits fast—especially if you’re staring down a legacy codebase that has developed the personality of a haunted house.

The hype machine is fueled by three things that are very real:

  • The tools are genuinely useful for certain classes of work (boilerplate, translation between languages, test scaffolding, docs, common patterns).
  • They’re improving quickly and product teams ship new features at the pace of “surprise, it’s Tuesday.”
  • They’re easy to measure badly. If you measure productivity as “lines of code produced,” you will always be impressed—right up until you have to maintain those lines.

But the “AI will write 90% of code by next quarter” narrative tends to skip the boring parts of software engineering: understanding requirements, dealing with edge cases, integrating with existing architecture, coordinating across teams, and—my personal favorite—debugging issues that only happen on the CEO’s iPhone while he’s in an elevator.

The evidence: productivity gains are real… and not universal

To cut through AI coding hype, the most useful move is to read studies that are not marketing material, not benchmark leaderboards, and not your friend’s viral thread about “I built a SaaS in 45 minutes.” Let’s look at the research signals that matter.

Field experiments: AI can slow down experienced developers (yes, really)

A particularly bracing data point comes from Model Evaluation & Threat Research (METR). In a randomized controlled trial of experienced open-source developers working on their own repositories, METR found that when developers were allowed to use early-2025 AI tools, they took 19% longer to complete tasks than when they weren’t allowed to use AI. Even more awkward: developers believed they were faster. METR reported a perception gap where developers expected a ~24% speedup and still believed they were ~20% faster after the work—despite actually being slower. METR’s write-up is here, and Ars Technica’s coverage is here. citeturn7search2turn7search0

This doesn’t mean “AI never helps.” METR explicitly warns against overgeneralizing. It does mean two things that engineering leaders should tattoo onto their sprint boards:

  • Tool value depends on context. On big, old, familiar codebases with implicit standards and tribal knowledge, AI can create overhead.
  • Humans are bad at self-measuring productivity. Feeling faster is not the same as being faster.

Controlled studies: time savings often show up in the “boring” tasks

Other empirical work, including studies on GitHub Copilot usage, frequently finds improvements—especially in documentation, autocompletion, and repetitive coding tasks. One 2024 arXiv study evaluating Copilot in real-world projects reported significant reductions in “developer toil” for certain task types, while also noting struggles with complex, multi-file, proprietary-context work. citeturn1academia13

That last clause matters. Most organizations don’t just need “more code.” They need correct code that fits a system. AI is more like a junior developer who never sleeps and has read every blog post on the internet—including the bad ones.

Benchmarks: SWE-bench Verified shows progress, but it’s not your codebase

Benchmarks are useful, but only if you understand what they measure. OpenAI’s SWE-bench Verified is a human-validated subset of SWE-bench intended to more reliably evaluate whether models can solve real issues in GitHub repositories. OpenAI introduced it in August 2024 and updated the post in February 2025. citeturn8search0

Meanwhile, the public SWE-bench leaderboard has turned into a spectator sport—useful for tracking capability trends, but still a simplified stand-in for real production constraints. citeturn8search2

Benchmarks can tell you “agents are getting better at fixing issues in curated repos.” They can’t tell you whether an agent will understand your internal billing logic, your compliance constraints, or why that one service is named new_new_final_v3.

Why AI coding hype persists: it’s not (only) dishonesty, it’s incentives

AI coding hype isn’t just vendor overreach. It’s also the natural outcome of incentives across the ecosystem:

  • Vendors need growth and differentiation in a crowded market.
  • Enterprises want cost savings, talent leverage, and a story for the board.
  • Developers want relief from tedious tasks (and, frankly, to keep up).
  • Investors want “AI-native” narratives that justify valuations.

All of those incentives reward simple, dramatic claims. “We saved 30 minutes writing tests” doesn’t trend; “We fired half the dev team” trends, even when it’s an operational faceplant waiting to happen.

What AI coding tools are actually good at in 2026

Based on the research, surveys, and product direction, a pragmatic view looks like this:

1) Drafting and accelerating the first 60%

AI is often strongest at producing a plausible first draft: a function skeleton, a REST client, a UI component, an infrastructure template. That’s useful because the first draft is where humans tend to procrastinate. But it’s also dangerous because the first draft is where subtle design errors get baked in.

2) Documentation and explanation (with verification)

Tools are increasingly used for documentation and code explanation. That can pay off immediately for onboarding and knowledge sharing—but you still need verification, because “sounds right” is not a proof.

3) Test scaffolding (with a human in the loop)

AI can generate unit test scaffolds and happy-path cases quickly. The missing piece is usually domain-specific edge cases and meaningful assertions. In other words: AI can fill the spreadsheet; you still need to know the business.

4) Translation work

Porting between languages, frameworks, and API versions is tedious. AI is often quite good at this—again, provided you validate correctness, performance, and security.

The less-fun reality: AI coding changes your risk profile

Even if AI makes your developers faster, it can also make you faster at shipping bugs. Speed is a neutral force. The question is: what are you accelerating?

Security risks: context poisoning and agent exploits

As tools become more “agentic” (they gather context, run commands, open PRs), the attack surface grows. Academic work has shown how attackers can manipulate context sources to influence assistant output—like cross-origin context poisoning attacks against coding assistants. citeturn3academia15

Other research has analyzed real-world coding agents and found security issues that can enable command execution or data exfiltration under certain conditions. citeturn3academia17

This matters because organizations are increasingly giving these tools access to repositories, build systems, and internal documentation—effectively putting an eager intern in front of your crown jewels, then telling it to “be autonomous.”

AI-generated code vulnerabilities are measurable (and not rare)

Industry research keeps flagging that AI-generated code can be functionally correct and still insecure. For example, Veracode’s 2025 GenAI Code Security Report (announced via Business Wire) reported that AI-generated code introduced security vulnerabilities in a significant portion of evaluated tasks. citeturn3search1

Separately, a 2025 arXiv large-scale analysis examined public GitHub repositories with code explicitly attributed to AI tools and identified thousands of CWE instances via CodeQL across many vulnerability types. citeturn3academia20

The takeaway is not “ban AI.” It’s “treat AI output as untrusted input.” Same as you would treat code from a random gist—because, statistically, it might as well be.

Behavior risk: developers don’t always check AI code

There’s also a human-factor problem. A Sonar survey discussed in recent coverage found many developers don’t fully trust AI-generated code—yet not all consistently review it before committing. That combination (low trust + inconsistent verification) is the sort of thing that keeps security teams awake. citeturn3news12

A practical framework: how to “cut through AI coding hype” on your team

If you lead engineering, you don’t need a philosophical position on AI. You need a policy that reduces risk while preserving upside. Here’s a pragmatic approach.

1) Measure outcomes, not vibes

  • Cycle time (issue opened → merged) instead of “time typing.”
  • Rework rate (how much code is rewritten/deleted within weeks).
  • Incidents and regressions tied to AI-assisted changes.
  • Security findings (SAST/DAST/CodeQL deltas) on AI-heavy PRs.

2) Use AI where it’s strongest

Start with:

  • docs + comments
  • test scaffolding
  • small refactors
  • internal tooling
  • code search/explanation (with citations if possible)

Be cautious with:

  • auth and identity flows
  • cryptography
  • financial calculations
  • privacy-sensitive data handling
  • deployment automation with broad permissions

3) Put guardrails in the pipeline

  • Require code review (yes, still) and encourage “explain this diff” norms.
  • Turn on security scanning by default (CodeQL/SAST, dependency scanning, secret scanning).
  • Use policy controls for what code can be sent to external models (or use approved enterprise deployments).
  • Log and audit agent actions if you’re using autonomous workflows.

4) Train developers to prompt like engineers, not like marketers

The best prompts read like mini-specs:

  • Define inputs/outputs
  • List constraints (performance, security, backward compatibility)
  • Ask for tests
  • Ask for a threat model
  • Ask for trade-offs and alternatives

“Make it work” is how you get code that works until Tuesday.

Biotech trends to watch in 2026: AI is the new lab equipment

The second half of the MIT Technology Review newsletter title points at biotech trends worth watching. The biotech story of 2026 is increasingly a compute story: more data, more automation, more AI, and more pressure to prove real-world outcomes.

Here are the themes that are most likely to shape biotech strategy (and tech strategy inside biotech) this year.

Trend 1: AI-native drug discovery moves from “demo” to “industrialization”

Drug discovery has always been an information problem wrapped in a chemistry problem wrapped in a biology problem. The pitch for AI is that it can compress the search space—finding targets, generating candidate molecules, predicting properties, and triaging what’s worth synthesizing.

What’s changed is the scale of investment and the operational approach. At the 2026 JPMorgan Healthcare Conference, Nvidia announced partnerships with Eli Lilly and Thermo Fisher Scientific aimed at building AI-powered infrastructure for drug discovery and autonomous labs. Nvidia said the collaboration with Lilly would involve an AI innovation lab built on Nvidia’s BioNeMo platform and Vera Rubin hardware, with up to $1 billion in investment over five years. citeturn2news12

This matters because it signals a shift from “AI helps scientists” to “AI helps run the pipeline.” When companies start spending like this, they’re not buying toys—they’re buying an operating model.

Trend 2: Autonomous labs and the rise of the ‘robotic wet lab stack’

Biotech is taking cues from cloud computing: standardize, automate, observe, iterate. Autonomous labs—robotics plus software plus AI planning—promise faster experiments, better reproducibility, and tighter feedback loops between hypothesis and validation.

But the industry will also learn (again) that automation doesn’t eliminate complexity; it moves it. Your lab can be fully robotic and still fail because your data schema is a mess or because someone labeled tubes like it was 2009.

Trend 3: Cloud and data platforms become core biotech infrastructure

Between multi-omics data, imaging, EHR-linked studies, and increasingly AI-heavy analytics, biotech organizations are leaning hard on cloud and high-performance compute to scale. Several industry trend pieces highlight cloud adoption and real-time analytics as key enablers for faster R&D cycles and collaboration. citeturn0search3turn0search5

If you work in software, this is familiar: the winners will be the companies that treat data engineering as a first-class product, not a side project for “when things calm down.” Things do not calm down.

Trend 4: Regulation catches up to AI-enabled medical devices

Biotech and health tech in 2026 aren’t just about molecules—they’re also about AI-enabled devices and diagnostics. Regulators are increasingly focused on the total product lifecycle: development, validation, updates, monitoring, and risk management after deployment.

The U.S. FDA issued a comprehensive draft guidance for developers of AI-enabled medical devices on January 6, 2025, framing expectations across the device’s Total Product Life Cycle (TPLC). The FDA also noted it had authorized more than 1,000 AI-enabled devices through established premarket pathways. citeturn2search0turn2search2

There’s also active work around postmarket monitoring methods—detecting data drift, monitoring performance, and evaluating real-world behavior. citeturn2search1turn2search3

For biotech leaders, this implies two operational needs:

  • MLOps for compliance: versioning, traceability, monitoring, and change control.
  • Security by design: because connected health systems and ML components expand the threat surface, as research has highlighted. citeturn2academia14

Trend 5: Multi-omics and liquid biopsy diagnostics keep scaling

Diagnostics are trending toward earlier detection and broader panels—often combining multiple signal types (DNA, RNA, proteins, methylation) and using AI to interpret them. Industry trend roundups point to liquid biopsy and multi-omics diagnostics progressing toward broader screening and more connected workflows, alongside discussions about follow-up care and health equity. citeturn0search1

The technical challenge is not just building a model—it’s building a clinical pathway that makes sense when the model fires. “We detected a weak signal” is not a care plan.

Where AI coding and biotech collide: software is now a lab instrument

Here’s the connective tissue between the two halves of the newsletter topic: biotech is becoming software-driven at a deeper level, and software engineering is being reshaped by AI. That creates a new shared set of challenges:

  • Reproducibility: In biology, irreproducible results are costly. In software, irreproducible builds are… also costly. AI adds another layer of variability unless you lock down workflows.
  • Traceability: Regulators want audit trails. AI coding tools can generate code quickly, but you still need to know why decisions were made.
  • Security: Both domains are high-value targets. AI-assisted pipelines can become new attack surfaces.
  • Talent: The most valuable people are “bilingual”: they understand both the domain (biology or medicine) and the systems (data, cloud, ML, and now AI-assisted development).

In other words: the biotech trends to watch are also DevOps trends to watch—just with pipettes.

Actionable takeaways (because you still have meetings)

If you’re an engineering leader adopting AI coding tools

  • Run a 6–8 week pilot with real metrics (cycle time, rework, incidents, security findings).
  • Document allowed and disallowed use cases and revisit quarterly—tools change fast.
  • Assume AI output is untrusted and enforce scanning + review.
  • Invest in enablement: prompt training, examples, and shared best practices.

If you’re in biotech (or building software for biotech)

  • Prioritize data foundations (governance, lineage, interoperability) before scaling AI bets.
  • Design for monitoring and drift from day one, especially for clinical-facing systems.
  • Expect regulators to ask lifecycle questions, not just “does it work in validation?”
  • Plan security early—connected systems plus ML equals expanded attack surface.

Closing thoughts

The most useful way to think about AI coding in 2026 is not “replacement,” but “leverage.” It can absolutely reduce toil and accelerate certain tasks. But the more you treat it like magic, the more likely it is to hand you a magic trick: something that looks impressive until you examine it closely.

And on the biotech side, the most important trend is that biology is becoming compute-native. AI isn’t just an add-on; it’s becoming part of how experiments are designed, executed, and interpreted. That’s exciting—and it will reward organizations that pair ambition with operational discipline.

So yes: cut through the hype. Keep the useful parts. Throw the rest into the same drawer as 3D TV, blockchain toothbrushes, and the “metaverse strategy deck” nobody admits they wrote.

Sources

Bas Dorland, Technology Journalist & Founder of dorland.org