The internal data shared with The Economist do not identify spending by country, but grant titles suggest that South Africa has been hit particularly hard. Between 2022 and 2024 a yearly average of 18 new foreign grants contained South Africa in their title. In 2026 that figure dropped to one. A separate analysis by Grant Witness, a non-profit, suggests that the focus of funded research is also changing. A comparison of grants shows that references to Africa had been removed from 184 grant abstracts in 2025 and 2026 (see chart 3). Researchers in South Africa have also been tangled up in Mr Trump’s dispute with their government over allegations of a “white genocide” against Afrikaners. In February 2025 Mr Trump halted American aid to South Africa, a restriction the NIH interpreted as applying to its research grants, too. It was not until August 2026 that Dr Bhattacharya announced that NIH grants were exempt, though that came too late for many projects, like Dr Joseph Davey’s. The damage to international research networks is so far more tangible than the improvements to oversight. Under the previous system much of the quotidian monitoring of foreign grants fell to American universities and their armies of compliance officers, lawyers and contractors. A former university vice president for research, speaking on the condition of anonymity, acknowledged that institutions’ diligence varied enormously. A better
approach, they argue, would have been to demand more rigour from all universities and hold them accountable when they fell short. Instead the NIH has stripped universities of their primary means of enforcing compliance: control of the purse. Dr Bhattacharya says that foreign institutions are “building up their capacity to comply with American oversight requirements”. The change, he argues, will allow NIH professionals to take the oversight burden away from universities. How exactly this would be done remains unclear. “The NIH does not have a detailed investigatory culture or enforcement capacities,” says Doni Bloomfield, a law professor at Fordham University in New York. Meanwhile, staff across the NIH argue that the new system has increased paperwork while incentivising applicants to find loopholes. “This is all just making people minimise, recharacterise or completely hide foreign involvement,” says one programme officer. “We are going to get dramatically less oversight.” ■ This article was downloaded by zlibrary from https://www.economist.com/science-and-technology/2026/11/18/america-has-slashed- funding-for-biomedical-science-abroad
Science & technology | Caught on the hop Some grasshoppers fly like butterflies to mislead predators It is a rare example of motion-related mimicry Nov 19th 2026 FOR A SMALL animal in a world full of predators, looking good enough to eat is a fatal mistake. Looking like something disgusting or dangerous, though, is a valuable skill. It is called Batesian mimicry. Ever since the naturalist Henry Bates described the concept in 1862, it has been found everywhere. There are flies that look like wasps, birds that look like toxic caterpillars and harmless caterpillars that look like snakes. New work led by Tom Sherratt, an entomologist at Carleton University in Ottawa, is revealing evidence of a rare form of Batesian mimicry: the Carolina grasshopper (Dissosteira carolina) makes potential predators believe that pursuit would be a waste of energy.
Dr Sherratt’s interest in the grasshopper began when he moved to Canada from Britain and was told that the butterflies he kept seeing were actually grasshoppers. He discovered that collectors often nab grasshoppers by mistake when trying to capture the butterfly known as the mourning cloak (Nymphalis antiopa), which is common in North America. Dr Sherratt suspected Batesian mimicry. The imitation was clear: both animals share black-and-yellow wings. The benefit was harder to pin down. The butterfly is not unpalatable, poisonous or venomous. Why mimic it? Dr Sherratt and his team compared the colours of the insects as well as the “lazy flaps” of the Carolina grasshopper’s wings, a feature which often fooled butterfly collectors. The researchers also took high-speed videos of the bugs in the lab and in the field, and of other grasshoppers from the same region. The team’s analysis revealed modest similarities in colour between the two animals, but not the close match often seen in other Batesian mimics. The yellow of the grasshopper wings was not quite the same hue of yellow as found on the butterflies. The grasshopper wings also lacked distinctive blue spots. In motion, however, the similarities become remarkable. Dr Sherratt and his team report in Evolutionary Ecology that the Carolina grasshopper beat its wings at least 26% more slowly than any of the other local grasshopper species. This brought its wingbeat frequency down from 40-50 hertz to around 20 hertz; not far off the mourning cloak butterfly’s average frequency of 10 hertz. The researchers also noticed that the grasshoppers spent much more time gliding around with wings outstretched than their peers did, thereby behaving almost identically to the mourning cloak. The grasshopper’s deceptive movement probably compensates for its feebler colour match, notes Dr Sherratt. There are two reasons, suspects Dr Sherratt, why the grasshopper has been driven by evolution to adapt in these ways. First, the mourning cloak is notoriously difficult to catch. It is evasive in the air and birds often ignore it
because the meal is not worth the pursuit. Given that grasshoppers are a staple of many bird diets, the Carolina variety benefits from its disguise. Second, the grasshopper’s in-flight resemblance to a yellow butterfly brings another (non-Batesian) benefit. When it lands, it abruptly folds away its wings to reveal a dirt-coloured body. Dr Sherratt suspects that the sudden change may trick a bird into thinking it has lost its prey and calling off the hunt. Further work is needed to confirm Dr Sherratt’s hypothesis. If he is right, it would prove what many theatrical impresarios already know: that lacklustre costuming can be saved by sufficiently spectacular choreography. ■ This article was downloaded by zlibrary from https://www.economist.com/science-and-technology/2026/11/18/some-grasshoppers-fly- like-butterflies-to-mislead-predators
Science & technology | Ups and downs SpaceX achieves another historic first It has caught the second stage of its Starship rocket Nov 19th 2026 IT WAS NOT the first time SpaceX has plucked a rocket from mid-air. But it was the first time it had caught the top half. On November 14th, during the 15th test flight of its giant Starship rocket, Elon Musk’s rockets-and-AI firm for the first time managed to return the second stage safely back to its launch tower. That leaves the company on the cusp of realising its long-standing goal of building the first-ever fully reusable rocket, something that should slash the cost of getting things into orbit. The flight was not a total success. SpaceX had hoped to achieve the double, catching both the rocket’s booster stage a few minutes after launch and its second stage hours later. In the event, the booster ran out of fuel, diverting itself over the ocean near SpaceX’s launch site on the Texas coast, before self-destructing. But the second stage performed flawlessly. It carried out its
characteristic belly-flop manoeuvre before carefully and gently flying itself back to the launch tower, where it was caught by a pair of arms. Recovering a rocket’s second stage—which must survive the fiery heat of re-entry—is much harder than recovering its booster, which does not travel very far from its launch site. SpaceX’s engineers will be particularly keen to examine Starship’s heat shield, which is made up of thousands of heat- resistant tiles. The tiles have proved finicky on previous flights, with some falling off during re-entry, or allowing hot plasma in through the gaps. Building a durable heat shield that can reliably survive many trips back through the atmosphere is vital to SpaceX’s goal of rapidly reusing individual Starships. The firm’s engineers already have some data: on a test flight on July 24th the second stage unexpectedly survived a planned ocean landing. But having an intact and dry second stage back at SpaceX’s home base—as opposed to one that landed far from home and spent weeks floating in the ocean—will nonetheless be invaluable. Starship’s enormous capacity is vital for SpaceX’s plans to expand its Starlink satellite-broadband system and to eventually put AI data centres into orbit. The firm hopes that, once the rocket is fully developed, it will be able to launch 200 tonnes of payload into low-Earth orbit. Mr Musk aspires to costs of less than $10m per launch. Even if the firm falls far short of that target, it will still have a rocket that is far cheaper and more capable than the Falcon 9, which is thought to cost $15m-20m per flight—the precise figure depends on which analyst you ask—to take 25 tonnes of cargo into orbit. SpaceX foresees enormous demand for Starship’s services, almost all of it internal. In August the firm bought 50,000 hectares of land on the Louisiana coast, upon which it plans to build at least ten launch pads in the hope of eventually launching dozens of flights a day. In the meantime it is looking to stop flying the Falcon 9 within the next two years. When—or if—SpaceX will offer Starship to external customers is not yet clear, leaving satellite- makers fretting about a looming launch shortage. The successful catch will have left SpaceX’s engineers delighted—and its external customers even more nervous than before.■ This article was downloaded by zlibrary from https://www.economist.com/science-and-technology/2026/11/15/spacex-achieves- another-historic-first
Science & technology | Bad language models LLMs are developing filthy mouths More challenging tasks may be to blame Nov 19th 2026 RARE IS THE soul who can face a trying task—assembling a bookcase, say, or writing an opening paragraph—without resorting to a little quiet profanity. For better or for worse, the use of bad language is a thoroughly human trait. And, it turns out, an inhuman one, too. According to number-crunching from Vals, an AI analyst firm, large language models (LLMs) trained after July 2026 swore more readily than their predecessors. “Models swear when stuck, and mostly to themselves,” said Connor Frank, the report’s author. Bad language in times of stress exists in the underlying training data, but tends to be hammered out of most systems by the “reinforcement learning” steps that see results ranked by human feedback. Mr Frank doesn’t fully