Immunology and Cell Biology, Год журнала: 2023, Номер 101(10), С. 923 - 935

Опубликована: Сен. 18, 2023

The emergence of large language models (LLMs) and assisted artificial intelligence (AI) technologies have revolutionized the way in which we interact with technology. A recent symposium at Walter Eliza Hall Institute explored current practical applications LLMs medical research canvassed emerging ethical, legal social implications for use AI-assisted sciences. This paper provides an overview symposium's key themes discussions delivered by diverse speakers, including early career researchers, group leaders, educators policy-makers highlighting opportunities challenges that lie ahead scientific researchers as continue to explore potential this cutting-edge such ChatGPT Bard These publicly available can generate cogent, human-like human-level responses a range across knowledge areas, education. However, advancements comes new set implications. Medical education are no exceptions, our organizations must contend governance responsibility. Chat-GPT Research (WEHI)1 was led lab heads, policy-makers, representing academic landscape within institutes who engage their work, experts fields learning navigate appropriate, efficient ethical application AI LLMs. Together, speakers sought provoke 500+ in-person online audience on research, its overtly friendly editor papers grants, ability turn non-coders into bioinformaticians, analyze big data warp speed. In addition AI-driven tools Alphafold protein hallucination, addressed broader societal using science, concerns around ethics, privacy, confidentiality security writing is entered ether. WEHI discussions, “Artificial intelligence” has certainly captured popular imagination since launch ChatGPT, but been used widely clinical some time.2 Models drive cars, recognize images even create synthetic – different several ways. As LLM, it immediately accessible, interacting model easy having conversation. Designed allow users enter natural “prompts” “generate” response, depending nature prompt, output often surpass knowledge, expertise efficiency human entering it. relevant many “human”-driven tasks from basic editing distillation topic complex analysis collation dispersed information. Crossing threshold science-fiction reality required significant technological financial investment development training neural network-based well-resourced technology focused companies collaborations. GPT4, example, Microsoft enable OpenAI startup evolution form. highlights scale “training” produce predictive text generating important logistical considerations implementation prior public release, how may considerations. made splash Language Network space before. earlier GPT2 LLM completion network launched 2019 had “limited” release amid full version be fake news articles or other nefarious purpose.3 1.5B parameter released soon after, when claimed fears turned out overestimation network's performance did not traverse “uncanny valley” (Figure 1). leap between subsequent iterations GPT3.5 most recently GPT4 stark. It interesting while given access model, they built safeguards mitigate malicious use, anyone almost come phrase “As I can-not…”. Artificial designed seen growth adoption years. created address various faced helping them streamline improve enhance quality research. There now exists expansive toolbox support inbuilt reference managers, image video analysis, survey experimental design platforms well plagiarism detectors (Table related landscape, making easier tackle focus more creative aspects work. While offer tremendous benefits, essential understand limitations biases these ensure reliability validity findings. Zotero Mendeley EndNote ChatPDF Scholarcy Explainpaper IBM SPPS R Pandas NumPy Google translate NLTK spaCy OpenCV TensorFlow DALL-E2 Cariyon teams Slack Workspace Elicit Qualtrics SurveyMonkey Semantic Scholar Iris.ai rabbit Discovery Tableau Power BI Turnitin IThenticate Copyscape identified two broad, overlapping considered adopting AI-based research: (1) wide communication confuse; (2) implications, future developments domains, law, intelligence, analyzed context what will mean scientist future. We discuss themes, consideration further impact science domains become integrated word-processing, spreadsheet multimedia software. following thought invoke among readers broad scientists, ethicists navigating limited, growing, understanding experience. Expert reviews found Table 2. Large applied breadth limited work performed bench, bridge barriers science. Indeed, there exciting accessibility facilitate collaboration non-native English speaker scientists parts world. real bias accuracy generation, particularly where objectivity paramount. Scientific literature vital advancing poor readability poses challenge. issue goes beyond technical jargon incorrect syntax. Common comprehension include excessive passive voice, long convoluted sentences unnecessarily language. Poorly written hinder effective impede dissemination findings community beyond. Additionally, increasingly competitive funding, convey significance sense excitement, remaining accessible. regard, emerged assistant. clarify ambiguous statements reader's understanding. also identify simplify terminology, accessible non-experts. beneficial grant writing, reviewers career-defining decisions lack subject matter expertise. pitfalls generative undermine ChatGPT's simplifying summarizing writing. statistically reconstructed does guarantee coherence. summaries encompassing questions, main findings, methodology, results unreliable inaccurate. Thus, expert author critically assess AI-generated through careful fact-checking cross-referencing. Even able inaccuracies domain, confidence misinformation present risk. model's information date restriction, excluding up-to-date unless newer link internet used. Nonetheless, appropriately leveraging capabilities, optimize time expertise, allowing improving help world move towards globalization international collaboration, importance proficient skills biomedical cannot overstated. Non-native face vocabulary, grammar rules cultural nuances, creating separation colleagues collaborators. context, empowering inclusive tool. accurately multiple languages, those alphabets 2). helps breaking down primary thousands languages spoken world, thus collaborations networks, globally. Excellent attribute being successful understands individuals secondary disadvantaged. Communication foundation productive environment collaborations, compromised unfamiliar styles cultures, experienced email practices. Here, gap. AI-aided translation emails collaborators vastly shorten needed formulate matters that, comparison, easily generated native 3). No longer restricted only translation, serve personal assistant, teacher translator, all one platform. Specialized text-to-voice software pronunciation, Kick Resume aid resume valuable resource content-based retrieve about specific biotechnological techniques, CRISPR, Additional prompts request corresponding references accuracy. excels proofreading With planning integrate co-pilot intending extend Suite productivity tools, inevitably day-to-day tasks. integrations assistance activities drafting emails, presentations interpreting content, stand benefit greatly incorporation tools. Although limitations, indeed translated requires verification, professional's powerful skills, individual foster inclusion community. Supervisors graduate students provide rounds feedback construct thesis. reviewing correcting structure form part thesis revision. courses professional copyediting services proofread thesis, universally cost-effective. typing assistants, Grammarly, already little controversy real-time spelling, grammar, punctuation clarity, suggesting replacements errors. Windows365 suite, just another icon, located next assistance. principle, should reduce load supervisors. Much less spent removing commas up paragraphs, talking Yet, raises sector regarding originality content produced student. Many supervisors worried lose critical so passionate developing students. need reflect own red underlined typos blue grammatical errors currently highlighted Word Powerpoint documents, don't simply accept them; review correct appropriate. any supervisor, copyeditor still revised student themself. acknowledge if appropriately, learn same traditionally have, differently, ok. unable undertake logical thinking true sense, assistance, search engine automates exceptional researchers. researcher asked organize conference Australian Consortium whose members interests readily websites publications. To richness, prompt initially included details title/theme, purpose, topics, target number, location, duration total budget. aim see whether could than generic program covering organization content. Unfortunately, result rather disappointing. Firstly, ten named organizing committee eight connection consortium theme. individuals, affiliations were incorrect. None known members. Attempts selection contributors improved regenerating additional keywords suggested speakers; fact, except copying declined. Secondly, align would traceable. Presentation titles too e.g. ‘Multiomics research’, ‘Controversies’, ‘Future Directions’. Finally, although provided sessions keynote oral poster presentations, workshops budget breakdown, closest got accurate suggest venues registration websites. At time, useful checklist conference, lacks go craft meaningful size complexity datasets grown over demand programming outpaced availability frequently leading bottleneck iteration. Analyzing daunting bench scientists. relying bioinformaticians analyses introduce delays challenges. Tools explicitly trained code, empower datasets. non-coding user describe inputs desired outputs code bioinformatic analysis. improvement evident reduced number refine execution task, needing quarter compared GPT3.5. study 97.3% bioinformatics task solved 7 prompts. Despite excellent results, points lead erroneous absence comprehension.8 huge prompting explain reasoning behind functions, resulted detailed description underlying algorithm. Furthermore, employed summarize interpret bioinformatician plain interpretation summary scientist, assessment script's limitations. Further script, comprehensible reusable applications. examine bioinformatics, controlled experiments classroom setting conducted.9 final might profession. expect immediate surge computational laboratories automate prototyping. medium term, like interpreter democratize enabling direct dataset via Python interpreter. skill likely shift syntactically better testing. envision liberated routine analyses, concentrate bespoke shift, turn, favor stronger logical, mathematical raw output. rise large-scale imaging multi-channel, multi-dimensional, long-term live-cell microscopy providing information-rich pipelines extract results. cope challenge quantifying data, few options: collaborate do (ideally specialist); try previously published pipelines; (3) rely proprietary modified. Incorporation workarounds always directly in-house datasets, techniques laboratories. Challenges achieving opening source pipeline detecting counting cells, clearly visible groups increase robustness model. New methods lattice light-sheet offering plug play (hours days), video-rate, 3D, answering biological questions system. modalities, handling ongoing afterthought. coding capability offers opportunity accelerate rapid generation packages. An example presented specialists assist framework least laying foundations workflow. (in case Python) import libraries, segment regions interest thresholding attempt quantify 4). then plotted accordingly. running light sheet quickly became apparent struggle overcome novel approaches For prompted track segmented cells plot paths detected cells. appreciate perfectly straight lines, LLM. Mathematical common, steps requiring quantification manual checking. upon fix errors, 2 3 potentially non-functional alternatives, coder's attention resolve. starting basis analytic pipelines, good researcher's clarity question and, importantly, user's fact-check verify GPT3.5, GPT4.0 major advances background. ChatGPT4.0 incorporated Omega (https://github.com/royerlab/napari-chatgpt) takes processing napari, attempts bugs real-time. ecosystem area becoming invaluable assistant acting alternative StackOverflow finding fixes concepts Theme 1 demonstrate real-world applications, effectiveness tempered assisting crucial guide edit appropriately. instances, find write themselves handle mundane time-consuming tasks, copyediting. advantages evident, whom second Opinions vary academics coding, largely depends compelling looking code; however, each step check valid treatment tool risky. wet extremely coders instances chunks multistep significantly slows process introducing require extensive debugging. Used judiciously discrete well-defined well-understood speed theme machine discussion revolved do. raised AI-assistance role intellectual contributions truly gained versus lost put profound working varying levels understanding, experience awareness shortcomings. continues shape continuously transformative actively strive responsibly us discoveries promote flourishing, needs aligned goals. Due inherent difficulty specifying undesired behaviors, sparked field called alignment core ensuring risks humanity kept minimum. harm humans ways; influencing commit unethical behavior enabler behavior. corrupting effects AI, behavioral based empirical observations. Adopting approach evidence-based policies. currently, primarily undertaken sell insist changing. Kobis et al.,7 evaluating human-computer interaction Whilst consensus enquiring minds discovery, mindful. shared queries pool train models. user, permits you sign up. restrictions placed risk once platforms, disclosure confidential. Companies Samsung banned after engineers inadvertently disclosing trade secret.10 Following incident, Walmart, Amazon implemented similar bans, until guidelines developed.11 universities established policies teaching, assessment. They remind value integrity, strong advocates (careful) Other posting onto platform, constitute breach contravene privacy laws. According policy, post bound General Data Protection Regulation, Processing Addendum provider. Questions copyright ownership moral rights' infringement Copyright subsists author. Since capable responds PhD Master's theses

Язык: Английский