Bard's latest updates: Access Gemini Pro globally and generate images

Original Article ↗ Hacker News Discussion ↗

Google’s latest upgrade to Bard, adding Gemini Pro access and image generation, is widely seen as underwhelming compared to OpenAI’s GPT‑4 and other competitors. Commenters highlight uneven availability (notably excluding Canada and image generation in much of Europe), heavy-handed censorship and safety filters, and inconsistent behavior that undermines real-world usefulness. Many see these issues as symptoms of Google’s regulatory caution and internal misalignment, raising questions about whether the company can catch up in large language models despite its early AI research lead.

Availability & “Global” Rollout

  • Many commenters note Bard is still unavailable in Canada, despite Google labeling access as “global.”
  • Image generation (Imagen 2) is also blocked in the EEA, Switzerland, and the UK per Google support, causing confusion for users who received feature announcements but can’t use them.
  • Explanations debated: Canadian bills (C‑18, C‑27), Quebec’s Bill 64, privacy commissioners investigating OpenAI, “small market” arguments, and general regulatory uncertainty.
  • Others counter that Bard runs in far stricter jurisdictions (EU), that OpenAI/Microsoft operate in Canada, and that Google hasn’t clearly articulated its reasoning.

Model Quality vs GPT-4 / Benchmarks

  • Multiple users report Bard/Gemini Pro underperforming GPT‑4 and often even GPT‑3.5 on reasoning, coding, and complex tasks; hallucinations and “bureaucratic” responses are common complaints.
  • Some find Bard better for creative writing or legal-style drafts, and note that the free GPT‑3.5 is “garbage” by comparison to paid GPT‑4.
  • Gemini Pro’s strong LMsys leaderboard position is viewed skeptically by some, who suspect benchmark overfitting or special variants; others attribute it to updated fine-tuning and internet access and warn against over-indexing on a single leaderboard.
  • Ultra (the model meant to compete directly with GPT‑4) is still not available, so some reserve judgment.

Image Generation & Safety Filters

  • Speed of Bard image generation is praised, but quality is often judged inferior to DALL‑E 3 and Midjourney, with limited stylistic range and trouble with complex compositions.
  • Safety filters are described as unusually aggressive and inconsistent: blocking bikinis, sadness, knives, certain racial descriptors, and sometimes any people or animals at all.
  • Users report shifting behavior over time (sometimes “can’t generate images,” sometimes only very constrained ones).
  • Google’s SynthID watermarking is noted; availability of public detection tools is unclear.

Censorship, Safety, and UX

  • Bard frequently refuses tasks (legal drafting, API code, translations with profanity, technical torque tables), citing safety, liability, or policy.
  • Many see this as heavier-handed than competitors; others note similar issues with Bing and GPT‑3.5.
  • Requests for transparent model-version indicators and less confusing rollouts (A/B testing, staged releases) are common.

Google’s Strategy & Competitive Position

  • Commenters debate whether Google’s lag comes from mismanagement, fear of disrupting its search/ads business, or greater legal/reputational risk tolerance.
  • Some predict long-term decline as users migrate to GPT‑4/Perplexity/Bing; others note Google’s scale, infrastructure, and cautious approach as rational but slower.