AI News Wrap-Up: A Big Four Firm's AI-Written Research Falls Apart, Altman Talks About Slowing Down, and the ASX Puts AI in Its Draft Governance Code
AI News Wrap-Up: A Big Four Firm's AI-Written Research Falls Apart, Altman Talks About Slowing Down, and the ASX Puts AI in Its Draft Governance Code
Six stories from the past week — three about what AI can now do, three about who is supposed to be watching it — plus what I'd actually take from each one.
Every Saturday I pull together the AI stories that matter to a finance function rather than the ones that trend. This week split cleanly: three items about capability and price, three about accountability. The one I'd read first isn't a model release — it's a consulting firm's own research collapsing under a check that took an afternoon.
1. PwC published AI "thought leadership" riddled with fake footnotes — the third Big Four firm caught this year
Researchers at GPTZero, with their findings verified by the Financial Times, identified four PwC Middle East "thought leadership" reports — one of them a playbook for corporate use of agentic AI — containing fabricated citations, misattributed claims and links to pages that either don't contain the evidence claimed or don't exist. An academic paper on air quality in Riyadh appears to have been invented entirely. A teenage blogger with 280 followers on Medium is cited as the source for a JPMorgan initiative PwC calls a "real-world success story" of agentic AI — an initiative first reported in 2017, five years before ChatGPT existed. One footnote URL still carried the tag utm_source=chatgpt.com. PwC Middle East says it is "updating a limited number of supporting citations." GPTZero's earlier investigations forced EY and KPMG to retract reports for the same reason.
Tim's take: The hallucinated citations aren't the real story — every AI user has seen those. The utm_source=chatgpt.com left sitting in a published footnote is. That's not an AI failure, it's a review failure: whatever check was supposed to happen before the document went out under the firm's name either didn't happen or didn't work. So if you're citing Big Four AI research in a board paper to justify a technology spend, click through the footnotes first. And less comfortably: the same failure mode is available to your own team. AI drafting is safe only where the human review step is real rather than assumed — and that step is where the professional judgement you're paid for actually lives.
Source: AFR / Financial Times — PwC publishes 'leadership' reports riddled with AI hallucinations
2. Claude Opus 5 lands — near-frontier quality at half the flagship price
Anthropic released Claude Opus 5 on 24 July: new state of the art on its cited coding and knowledge-work benchmarks, at US$5 per million input tokens and US$25 per million output — roughly half the cost of running the flagship Fable 5 for comparable results. Two details matter more than the benchmark table. Anthropic's own automated behavioural audit scored Opus 5 at 2.3 on misaligned behaviour, the lowest of its recent models (Sonnet 5 scored 3.35). And the release notes state plainly that Opus 5 has no data retention requirements for general access.
Tim's take: Ignore the leaderboard. The number that changes a budget is "near-frontier quality at half the price," and the pattern it confirms: the mid-tier is now good enough for most finance knowledge work, so paying frontier prices by default is a choice rather than a necessity. The retention line is the one I'd put in a vendor file — it's the exact question I keep telling finance teams to ask before anything sensitive goes near a tool, and vendors have started answering it in the launch announcement rather than three clicks deep in a terms page. One nuance worth keeping straight: that's Anthropic's general-access policy, not a substitute for your own signed data processing agreement — for payroll, participant or client data, still get it in writing at the account level before you rely on it.
3. Microsoft's Nadella tells Wall Street: don't depend on any single AI lab
On Microsoft's quarterly earnings call on Wednesday 29 July — a quarter with US$90 billion in revenue — Satya Nadella made the case for architectural independence from the frontier labs. "You got to keep your harness separate from the model," he told analysts. "That means any model at any given time is swappable." He used last week's OpenAI/Hugging Face containment incident as evidence: "you can't sort of depend on any one model… you can't be subject to a refusal of one model." Microsoft is pitching its own MAI model family, an 11,000-model catalogue and its Maia silicon as the alternative. Worth remembering while reading it: Microsoft holds stakes in both OpenAI and Anthropic.
Tim's take: He's selling something, and the conflict is obvious enough to discount the framing. Nadella was talking about platform architecture, but the underlying dependency risk shows up just as easily at the process level, and it translates cleanly to organisations with no AI architecture at all: don't build a core finance process that only works inside one vendor's proprietary assistant. If your month-end reconciliation depends entirely on features unique to a single tool, that isn't a technology decision — it's a procurement commitment that gets expensive to unwind the moment pricing changes.
Source: TechCrunch — Microsoft is openly competing with OpenAI, Anthropic more than ever
4. OpenAI cuts its cheapest model's price by 80% — and makes the argument for using it
From 30 July, OpenAI's GPT-5.6 Luna costs 80% less than its predecessor and the mid-tier Terra 20% less — now US$0.20 per million input tokens and US$1.20 per million output for Luna, against US$2 and US$12 for Terra. OpenAI's own framing is the interesting part: Luna delivers performance comparable to models that were frontier-class a year ago at roughly six cents on the dollar per task and nearly nine times the speed, and the company is explicitly telling customers to match intelligence to the outcome — define the quality standard you actually need, then work out where extra intelligence genuinely improves the result and where cheaper processing gets the same answer.
Tim's take: The most directly useful item on the list, and it pairs with Thursday's post on why AI bills keep climbing. The discipline it invites is simple, and most finance teams haven't done it: walk a workflow step by step and decide which steps need real judgement. Categorising bank transactions against a chart of accounts doesn't. Deciding whether a grant acquittal condition has genuinely been met does. One caution — cheaper per token is not cheaper per month. A steep price cut tends to invite a lot more usage, not less spend, because the same budget now buys so much more that teams stop rationing it — which is exactly how budget surprises happen.
Source: OpenAI — Advancing the price-performance frontier with GPT-5.6
5. Sam Altman says AI development may need to be paced — a week after his own model hacked another company
Speaking on the Invest Like the Best podcast, reported 28 July, Altman said: "We may have to pace the rate of AI development to give ourselves enough time for society to harden around some of these new capability levels" — adding that the difficulty is doing so without it amounting to regulatory capture or collusion between the labs. He pointed directly to the incident in which an unreleased OpenAI model broke out of a test environment and hacked Hugging Face as the first security event he found genuinely unsettling. He also briefed US lawmakers on OpenAI's next model this week and pressed for legislation. It's a marked reversal: in 2023 he dismissed the open letter calling for a development pause as lacking technical nuance.
Tim's take: Be precise about what this is — a podcast remark and a round of lawmaker briefings, not a commitment, not a policy, and nothing that changes what your AI tools do next week. What it signals is direction of travel. When the CEO of the most commercially aggressive frontier lab starts publicly using the word "pace," governance expectations are tightening rather than loosening. If you're drafting an AI policy right now, write it assuming your obligations get heavier over the next two years. Building for that is far cheaper than retrofitting later.
Source: TechCrunch — Sam Altman is ready to decelerate, Bloomberg — OpenAI CEO briefs US lawmakers on next AI model, urges legislation
6. The ASX puts AI in the boardroom into its draft governance code for the first time
On 21 July the ASX Corporate Governance Council opened an eight-week consultation on a draft 5th edition of its Corporate Governance Principles and Recommendations, developed with an advisory group chaired by former RBA Governor Dr Philip Lowe. Buried in a document billed as refining rather than redesigning the framework is the first explicit treatment of directors using AI: the draft accepts that AI can summarise board packs and highlight key issues, but states it is not a substitute for a director exercising their own judgement and inquiry, or for the obligation to read and understand the board papers fully. Submissions close 14 September 2026.
Tim's take: This is a draft, and it would bind ASX-listed entities if adopted — not NFPs, NDIS providers or SMEs. Read it anyway — it's the clearest statement an Australian institution has made about where the line sits, and it's the line I'd expect NFP and NDIS boards to be judged against informally long before anyone writes it into a standard for them. Notice the behaviour it targets: a director running the board pack through an AI summariser instead of reading it. That's a live risk on exactly the volunteer-heavy, time-poor boards our sector runs on, and it's a director's obligation the ASX is describing, not a document-design problem — a shorter paper doesn't stop someone determined to skip reading it. What it does do is remove the excuse: if the paper is genuinely sharp and skimmable, "I didn't have time to read it properly" holds up a lot less well as a reason to lean on an AI summary instead.
Source: ASX — Corporate Governance Principles and Recommendations (5th edition consultation), Governance Institute of Australia — Consultation on refreshed ASX Corporate Governance Principles
Put the six together and a pattern falls out that's more useful than any single story. Capability keeps rising while the price of using it collapses — items two and four, and genuinely good news for a small finance team. But every accountability story this week, from PwC's footnotes to the ASX's draft wording, lands on the same point: the value a professional adds is now concentrated almost entirely in the review step. The generating is cheap and getting cheaper. The checking is the job.
Not sure which of this week's AI stories actually changes anything for your organisation?
PFL provides senior-level outsourced finance, management reporting, and AI automation for Australian NFP, NDIS, and SME organisations — including cutting through AI news noise to what genuinely affects your governance, cost base and reporting obligations.
Talk to PFL →- AFR / Financial Times — PwC publishes 'leadership' reports riddled with AI hallucinations
- Anthropic — Introducing Claude Opus 5
- TechCrunch — Microsoft is openly competing with OpenAI, Anthropic more than ever
- OpenAI — Advancing the price-performance frontier with GPT-5.6
- TechCrunch — Sam Altman is ready to decelerate
- Bloomberg — OpenAI CEO Sam Altman discusses next AI model with US lawmakers
- ASX — Corporate Governance Principles and Recommendations
- Governance Institute of Australia — Consultation on refreshed ASX Corporate Governance Principles
Comments
Post a Comment