When Chatbots Meet the NICU: What Four New Studies Tell Families About Artificial Intelligence in Newborn Care

Studies by Xu (Frontiers in Artificial Intelligence, 2026), Huang (Translational Pediatrics, 2026), Liu (Frontiers in Artificial Intelligence, 2026) and Jiang (Digital Health, 2026), testing chatbots on resuscitation exams, home light-treatment questions, home oxygen leaflets, and newborn pain videos

Four 2026 studies tested chatbot-style artificial intelligence on newborn care tasks: exam questions, parents' questions about home light treatment for jaundice, leaflets about home oxygen, and videos of heel-prick tests. The tools often did well, sometimes remarkably so, but some made worrying mistakes. For families, the message is that chatbots may help explain information, while your baby's care team should guide every decision.

When a baby is in the neonatal intensive care unit (NICU), the hospital unit that looks after newborns who are premature or seriously ill, families must learn a new vocabulary almost overnight. Specialists use unfamiliar terms, and the printed handouts that come home are often written at a reading level harder than many people find comfortable [1]. So parents naturally reach for a phone, often late at night, and increasingly ask an artificial intelligence (AI) chatbot. Four research teams recently asked how well these tools really perform on newborn care topics [2][3][1][4].

How Families Found Answers Before Chatbots, and How the Chatbots Arrived

Traditionally, answers often came from people. A nurse at the bedside explained what the numbers on the monitor meant, a doctor described the plan during rounds, and a folded leaflet went home in the discharge folder. Later, search engines let families type a worry into a box and read whatever came back, from excellent hospital pages to unreliable forums.

Chatbots changed that. A large language model (LLM) is a computer program trained on enormous amounts of written text so that it can predict, word by word, what a helpful reply would look like. Instead of a list of links, it replies in smooth, friendly paragraphs, as if a knowledgeable person were typing back. That fluency is its charm and its risk, because an answer can sound right even when it is not.

Researchers first checked these programs the way we check human trainees: by giving them exams. Early versions of ChatGPT failed American Heart Association life support exams when they had a single attempt, scoring roughly between 64 and 76 percent, as reported by the authors of the first study [2][5]. Later versions scored higher on some exams: ChatGPT-4o scored 90.7% on the Pediatric Advanced Life Support exam [6] and 75.3% on the Neonatal Resuscitation Program exam over three attempts [7], again as reported by the authors of the first study [2]. Yet passing a test is not the same as helping a family. In one randomized trial (a study in which people are assigned by chance to different groups), ChatGPT-4 working alone outperformed physicians at diagnostic reasoning (working out what is wrong from a patient's story), but physicians given access to it did not improve significantly, as reported by the authors of the first study [2][8]. The Huang team describe an earlier randomized trial of 147 newborns finding that light-treatment failure did not differ significantly between home and hospital [3][9].

What the Four Studies Did in Everyday Terms

None of the four studies treated a baby or tested a treatment, so none is a clinical trial.

In the first study, researchers in China and Canada gave two chatbots, ChatGPT-5 and DeepSeek-R1, written questions on newborn resuscitation, the emergency steps that help a newborn start breathing [2]. The questions came from a hospital workshop in Shenzhen, the standard Neonatal Resuscitation Program textbook, and quizzes used in Edmonton, Canada. Each chatbot took the tests three times, a week apart, and its scores were set beside those of nurses, doctors, and other health professionals who had taken the same courses [2].

The second study imagined a parent at home. Seven chatbots were asked 15 caregiver questions about phototherapy, the light treatment for newborn jaundice, given at home for a baby whom doctors had already judged suitable for it [3]. The questions were built from the American Academy of Pediatrics guideline [10] and device manuals. Three experienced newborn specialists then graded 210 sets of answers without knowing which chatbot wrote them, scoring accuracy, completeness, consistency, warmth, and readability [3].

The third study looked at leaflets. Six chatbots were asked to write, simplify, and rewrite information for families about home oxygen for premature babies with chronic lung disease, and the results were compared with 20 human-written articles [1]. Two experienced NICU nurses graded all 360 texts without knowing the source [1].

The fourth study used video. Three AI models that can process moving images were shown short recordings of 426 heel-prick blood tests in 142 healthy newborns and asked whether the baby was in pain [4]. Answers were compared with ratings by trained nurses using the Neonatal Pain, Agitation, and Sedation Scale (N-PASS), a widely used checklist covering crying, behavior, facial expression, muscle tone, and vital signs [11]. Careful pain assessment matters because babies in the NICU undergo about 10 painful procedures per day on average [4].

What the Studies Found

On the exams, the chatbots held their own. On the Shenzhen workshop test, ChatGPT-5 scored 89.5% and DeepSeek-R1 92.1%, compared with 81.9% for health professionals before training and 91.7% after it [2]. On the textbook questions the chatbots scored 93.3% and 90.7%, and on the Canadian quizzes 95.2% and 81.0%, against 78.3% for professionals [2]. A score of 90% means about 9 questions right out of 10. The authors reported that the differences between chatbots and professionals were not statistically significant, meaning they could not rule out chance [2]. But they did worse on short written answers than on multiple-choice items: ChatGPT-5 got 97.0% of multiple-choice questions right but 66.7% of short-answer ones, and both chatbots stumbled on arithmetic, such as the amount of saline flush after adrenaline [2]. The professionals' scores came from earlier courses, so this was not a head-to-head contest. The authors cautioned that exam accuracy does not equal clinical skill [2].

For home light treatment, results varied a great deal. DeepSeek-R1 earned 19.31 points out of a possible 20, with about 88% of accuracy ratings judged completely correct, while the other six models scored between 15.00 and 16.73 [3]. On three safety-critical questions (when to stop treatment, side effects, and when to go to the hospital), DeepSeek-R1 earned acceptable ratings in all 18 of 18 ratings, but ChatGPT-4o did so in only 10 of 18 [3]. One model confused blue-light treatment for jaundice with a treatment for a skin condition, and every model invented some references and failed to tailor advice to a particular baby [3]. DeepSeek-R1 also scored highest for warmth, while another model used directive language such as "You must" [3].

For leaflets, the chatbot texts were rated as medically accurate as, or more accurate than, the human-written WeChat articles [1]. A surprise followed. When chatbots were told to use the simplest everyday words, the overall quality score fell for all six models, with no meaningful gain in how easy the text was to understand [1]. They did better when asked to rewrite an existing leaflet, which made it more understandable for all six models [1].

For the pain videos, the best model, Gemini-3-Pro, agreed with the nurses on 86.9% of videos, nearly 9 out of 10 [4]. It recognized nearly all pain videos but only about two out of three no-pain videos, so it leaned toward seeing pain [4]. Another model, ERNIE-4.5-Turbo, agreed 78.4% of the time and showed the reverse pattern, missing more pain [4].

What This Means for Families and Their Baby's Care

These studies suggest a middle path. The studies did not test these everyday uses directly, but a chatbot may help you understand a word you heard on rounds, turn a dense paragraph into plainer language, or help you organize questions for the care team. Chatbots can also sound confident while being wrong, invent references, and miss details about your own baby [3].

For that reason, decisions about treatment, medicine doses, when to stop light treatment, or what to do in an emergency belong with your baby's care team, and urgent concerns call for the unit's phone number or emergency services, not a chat window. Avoid typing your baby's name, birth date, hospital, or medical record details into general chatbots (a general precaution; the studies did not test privacy). Ask the NICU team which written materials, websites, or videos they recommend; their guidance fits your child and your hospital's equipment. If a chatbot explains something and you are unsure, bring the explanation to your nurse or doctor and ask whether it is right for your baby.

Finally, remember that no program watches your baby the way you do. Your observations and instincts matter, and you are always welcome to voice a concern, even when a screen says everything is fine.

What Researchers Are Working on Next

The resuscitation authors called for more research on whether chatbots could support training, with careful human oversight of tasks like dosing [2]. The home light team noted that its questions were simulated and measured no outcomes [3]. The leaflet team pointed out that no real caregivers were involved and that every AI-generated material needs professional review before reaching families [1]. The pain team used healthy babies from one hospital, called for broader testing, and saw such tools as assistants, not replacements for clinical judgment [4].

A Closing Thought for Families

These four studies offer a snapshot of where chatbots stand: impressive in some tasks, uneven in others, and always better with a person checking. Your baby's nurses, doctors, and above all you remain the experts on this child, and the best use of any new tool is as a way to ask better questions.

References

  1. Liu Z, Yang X, Zhang Y, et al. Evaluation of large language models in generating and optimizing educational materials for neonatal home oxygen therapy. Front Artif Intell. 2026;9:1770564. doi:10.3389/frai.2026.1770564 ↩
  2. Xu C, Chen Y, Skelding S, et al. Performance of large language models in neonatal resuscitation assessments versus healthcare providers: an exploratory study. Front Artif Intell. 2026;9:1838877. doi:10.3389/frai.2026.1838877 ↩
  3. Huang S, Huang Q, Hu W, et al. Comparative evaluation of seven large language models in providing home phototherapy care guidance for neonatal hyperbilirubinemia. Transl Pediatr. 2026;15(4):121. doi:10.21037/tp-2025-1-868 ↩
  4. Jiang X, Yang N, Shi T, et al. Deciphering the "non-verbal" code: a preliminary exploration of multimodal large language models for neonatal pain recognition. Digit Health. 2026;12:20552076261473729. doi:10.1177/20552076261473729 ↩
  5. Fijačko N, Gosak L, Štiglic G, et al. Can ChatGPT pass the life support exams without entering the American Heart Association course? Resuscitation. 2023;185:109732. doi:10.1016/j.resuscitation.2023.109732 ↩
  6. Kokulu K, Demirtaş MS, Sert ET, Mutlu H. ChatGPT and pediatric advanced life support: a performance evaluation. Resuscitation. 2024;205:110451. doi:10.1016/j.resuscitation.2024.110451 ↩
  7. Demirtas MS, Kokulu K, Tunc G. Evaluating artificial intelligence competency in education: performance evaluation of ChatGPT in the neonatal resuscitation program exam. Resuscitation. 2025;209:110563. doi:10.1016/j.resuscitation.2025.110563 ↩
  8. Goh E, Gallo R, Hom J, et al. Large language model influence on diagnostic reasoning: a randomized clinical trial. JAMA Netw Open. 2024;7(10):e2440969. doi:10.1001/jamanetworkopen.2024.40969 ↩
  9. Pettersson M, Eriksson M, Odlind A, Ohlin A. Home phototherapy of term neonates improves parental bonding and stress: findings from a randomised controlled trial. Acta Paediatr. 2022;111(4):760-766. doi:10.1111/apa.16231 ↩
  10. Kemper AR, Newman TB, Slaughter JL, et al. Clinical practice guideline revision: management of hyperbilirubinemia in the newborn infant 35 or more weeks of gestation. Pediatrics. 2022;150(3):e2022058859. doi:10.1542/peds.2022-058859 ↩
  11. Hummel P, Puchalski M, Creech SD, Weiss MG. Clinical reliability and validity of the N-PASS: neonatal pain, agitation and sedation scale with prolonged pain. J Perinatol. 2008;28(1):55-60. doi:10.1038/sj.jp.7211861 ↩