ABSTRACT CHAPTER 1: In this paper, we integrate insights from the Knowledge Spillover Theory of Entrepreneurship and Innovation (KSTE+I) with Schumpeter's concept of the entrepreneur as a “factor of disequilibrium”. Specifically, we examine whether there is a correlation between the level of Artificial Intelligence (AI) knowledge available in a region and the number of newly established innovative ventures, defined as startups that file patents in any technological field within the same year they are founded. Empirically, we test for 287 Nuts-2 European regions whether the local AI stock of knowledge exerts an enabling role in fostering innovative entry within AI-related local industries (AI technologies as focused enablers) and within non-AI-related local industries, as well (AI technologies as generalized enablers). Results from fixed-effect regressions using Poisson and Negative Binomial models - while controlling for a range of concurrent drivers of entrepreneurship - indicate that the local stock of AI knowledge fosters the proliferation of innovative startups within AI-related local industries. This finding supports both the KSTE+I framework and the enabling role of AI technologies; however, it does not support the notion of AI technologies as generalized enablers for the creation of new innovative startups. ABSTRACT CHAPTER 2: This paper introduces a deep learning methodology employing transformer-based models to systematically identify Artificial Intelligence (AI) patents. After carefully reviewing the existing literature on the methodologies employed to trace AI developments, we develop two domain-specific classifiers tailored to two foundational AI fields: Learning and Symbolic Systems (LS-SYS) and Robotics and Autonomous Systems (RA-SYS). Building on the BERT (Bidirectional Encoder Representation from Transformers) for Patents foundational model (Srebrovic & Yonamine, 2020), we propose a fine-tuning pipeline that unfolds in three stages. First, we derive two domain-specific lists of 221 weighted n-grams by mining the AI scientific literature, which are used to assemble seed sets for both domains extracted from the Patent Universe. Second, we apply the patent landscaping procedure (Abood and Feltenberger, 2018) to expand these seed sets and generate anti-seed examples via negative sampling. Third, we fine-tune each BERT model on the resulting training corpora, yielding classifiers that achieve F1 scores above 0.90, which are publicly available on the Hugging Face Hub. Our models reveal a sharp post-2012 “Deep Learning Revolution” inflection in AI patenting activity, particularly for the LS-SYS domain, alongside pronounced geographic concentration in the United States and Asia. Sectoral analyses reveal that ICT industries lead LS-SYS inventions, while a heavy industry core drives the RA-SYS domain, with collaboration networks and CPC-based maps providing a more nuanced picture. Firm-level rankings spotlight ICT and software incumbents (IBM, Microsoft, Google), robotics manufacturers (Fanuc, Yaskawa), automakers (Toyota, Honda, Ford), and agile newcomers (Waymo, Zoox, X Development). ABSTRACT CHAPTER 3: This study examines how regional technological relatedness and local AI knowledge influence regional inventive activity, as measured by patenting activity. Using a novel three-way longitudinal dataset (670 four-digit CPC classes × 302 NUTS-2 regions × nine four-year periods, 1986–2021), we show that two broad mechanisms operate in parallel. First, in accordance with the extant literature, technologies that are cognitively close to a region’s existing patent portfolio enjoy higher patenting activity, confirming that relatedness remains a strong and persistent predictor of inventive output. Second, local AI endowments are positively associated with patenting across technological fields, even after conditioning on relatedness, indicating that AI plays an enabling and cross-cutting role in a given regional innovation system. Moreover, the interaction between relatedness and AI turns out to be negative and statistically significant, implying that AI attenuates the extent to which local inventive activity depends on the technology’s proximity to the regional portfolio. In sum, AI appears to raise the overall local inventive activity and partially relaxes its dependence on relatedness.
ABSTRACT CAPITOLO 1: In questo articolo integriamo le intuizioni della Knowledge Spillover Theory of Entrepreneurship and Innovation (KSTE+I) con il concetto schumpeteriano dell’imprenditore come “fattore di disequilibrio”. In particolare, esaminiamo l'esistenza di una correlazione tra il livello di conoscenza in Intelligenza Artificiale (IA) disponibile in una regione e il numero di nuove imprese innovative costituite, definite come startup che depositano brevetti in qualsiasi campo tecnologico nello stesso anno in cui vengono fondate. Dal punto di vista empirico, analizziamo 287 regioni europee NUTS-2 per verificare se lo stock locale di conoscenza in IA eserciti un ruolo abilitante nel favorire l’ingresso innovativo sia nelle industrie locali correlate all’IA (tecnologie IA come abilitatori focalizzati), sia nelle industrie locali non correlate all’IA (tecnologie IA come abilitatori generalizzati). I risultati delle regressioni con effetti fissi, stimate tramite modelli Poisson e Negative Binomial — controllando per una serie di fattori concomitanti che influenzano l’imprenditorialità — indicano che lo stock locale di conoscenza in IA favorisce la proliferazione di startup innovative nelle industrie locali correlate all’IA. Questo risultato supporta sia il framework KSTE+I sia il ruolo abilitante dell’IA; tuttavia, non conferma l’ipotesi secondo cui le tecnologie legate all’IA agiscano come abilitatori generalizzati nella creazione di nuove startup innovative. ABSTRACT CAPITOLO 2: Questo articolo introduce una metodologia di deep learning basata su modelli transformer per identificare in modo sistematico i brevetti relativi all’Intelligenza Artificiale (IA). Dopo aver esaminato attentamente la letteratura esistente sulle metodologie utilizzate per tracciare gli sviluppi dell’IA, sviluppiamo due classificatori specifici per dominio, progettati per due ambiti fondamentali dell’IA: Learning and Symbolic Systems (LS-SYS) e Robotics and Autonomous Systems (RA-SYS). Basandoci sul modello BERT for Patents (Bidirectional Encoder Representations from Transformers) (Srebrovic & Yonamine, 2020), proponiamo una pipeline di fine-tuning articolata in tre fasi. In primo luogo, ricaviamo due liste specifiche per ciascun dominio, per un totale di 221 parole chiave pesate, ottenute attraverso l'impiego di tecniche di text mining applicate alla letteratura scientifica sull’IA. Queste liste, congiuntamente a specifiche euristiche, sono utilizzate per costruire insiemi di brevetti seed per entrambi i domini, estratti dall’universo dei brevetti. In secondo luogo, applichiamo la procedura di patent landscaping (Abood e Feltenberger, 2018) per espandere questi insiemi seed e generare esempi anti-seed tramite negative sampling. In terzo luogo, effettuiamo il fine-tuning di ciascun modello BERT sui corpora di addestramento risultanti, ottenendo classificatori che raggiungono F1 score superiori a 0.90, resi pubblicamente disponibili sull’Hugging Face Hub. I nostri modelli evidenziano una netta accelerazione dell’attività brevettuale legata all’IA dopo il 2012, associata alla cosiddetta “Deep Learning Revolution”, particolarmente pronunciata nel dominio LS-SYS, insieme a una marcata concentrazione geografica negli Stati Uniti e in Asia. Le analisi settoriali mostrano che le industrie ICT guidano le invenzioni LS-SYS, mentre un nucleo di industrie pesanti traina il dominio RA-SYS. L’analisi a livello di impresa mette in luce il ruolo di grandi incumbent nei settori ICT e software (IBM, Microsoft, Google), produttori di robotica (Fanuc, Yaskawa), case automobilistiche (Toyota, Honda, Ford) e nuovi entranti innovativi (Waymo, Zoox, X Development). ABSTRACT CAPITOLO 3: Questo studio analizza come la "relatedness" tecnologica regionale e la disponibilità locale di conoscenza legata all'Intelligenza Artificiale (IA) influenzino l’attività inventiva regionale, misurata tramite l’attività brevettuale. Utilizzando un nuovo dataset longitudinale tridimensionale (670 classi CPC × 302 regioni NUTS-2 × nove periodi quadriennali, 1986–2021), mostriamo che due meccanismi operano in parallelo. In primo luogo, in linea con la letteratura esistente, le tecnologie cognitivamente prossime al portafoglio brevettuale già presente in una regione presentano livelli più elevati di attività brevettuale, confermando che la relatedness tecnologica rappresenta una determinante persistente della produzione inventiva. In secondo luogo, le dotazioni locali di IA risultano positivamente associate all’attività brevettuale in diversi campi tecnologici, suggerendo che l’IA svolge un ruolo abilitante e trasversale all’interno dei sistemi regionali di innovazione. Inoltre, l’interazione tra relatedness e IA è negativa e statisticamente significativa, indicando che la presenza di conoscenze in IA attenua la dipendenza dell’attività inventiva locale dalla prossimità tecnologica rispetto al portafoglio regionale esistente. Nel complesso, questi risultati suggeriscono che l’IA rafforza l’attività inventiva regionale complessiva, contribuendo al tempo stesso a ridurre parzialmente il vincolo imposto dalla relatedness tecnologica.
ESSAYS ON THE ECONOMICS OF ARTIFICIAL INTELLIGENCE
D'Alessandro, Francesco
2026
Abstract
ABSTRACT CHAPTER 1: In this paper, we integrate insights from the Knowledge Spillover Theory of Entrepreneurship and Innovation (KSTE+I) with Schumpeter's concept of the entrepreneur as a “factor of disequilibrium”. Specifically, we examine whether there is a correlation between the level of Artificial Intelligence (AI) knowledge available in a region and the number of newly established innovative ventures, defined as startups that file patents in any technological field within the same year they are founded. Empirically, we test for 287 Nuts-2 European regions whether the local AI stock of knowledge exerts an enabling role in fostering innovative entry within AI-related local industries (AI technologies as focused enablers) and within non-AI-related local industries, as well (AI technologies as generalized enablers). Results from fixed-effect regressions using Poisson and Negative Binomial models - while controlling for a range of concurrent drivers of entrepreneurship - indicate that the local stock of AI knowledge fosters the proliferation of innovative startups within AI-related local industries. This finding supports both the KSTE+I framework and the enabling role of AI technologies; however, it does not support the notion of AI technologies as generalized enablers for the creation of new innovative startups. ABSTRACT CHAPTER 2: This paper introduces a deep learning methodology employing transformer-based models to systematically identify Artificial Intelligence (AI) patents. After carefully reviewing the existing literature on the methodologies employed to trace AI developments, we develop two domain-specific classifiers tailored to two foundational AI fields: Learning and Symbolic Systems (LS-SYS) and Robotics and Autonomous Systems (RA-SYS). Building on the BERT (Bidirectional Encoder Representation from Transformers) for Patents foundational model (Srebrovic & Yonamine, 2020), we propose a fine-tuning pipeline that unfolds in three stages. First, we derive two domain-specific lists of 221 weighted n-grams by mining the AI scientific literature, which are used to assemble seed sets for both domains extracted from the Patent Universe. Second, we apply the patent landscaping procedure (Abood and Feltenberger, 2018) to expand these seed sets and generate anti-seed examples via negative sampling. Third, we fine-tune each BERT model on the resulting training corpora, yielding classifiers that achieve F1 scores above 0.90, which are publicly available on the Hugging Face Hub. Our models reveal a sharp post-2012 “Deep Learning Revolution” inflection in AI patenting activity, particularly for the LS-SYS domain, alongside pronounced geographic concentration in the United States and Asia. Sectoral analyses reveal that ICT industries lead LS-SYS inventions, while a heavy industry core drives the RA-SYS domain, with collaboration networks and CPC-based maps providing a more nuanced picture. Firm-level rankings spotlight ICT and software incumbents (IBM, Microsoft, Google), robotics manufacturers (Fanuc, Yaskawa), automakers (Toyota, Honda, Ford), and agile newcomers (Waymo, Zoox, X Development). ABSTRACT CHAPTER 3: This study examines how regional technological relatedness and local AI knowledge influence regional inventive activity, as measured by patenting activity. Using a novel three-way longitudinal dataset (670 four-digit CPC classes × 302 NUTS-2 regions × nine four-year periods, 1986–2021), we show that two broad mechanisms operate in parallel. First, in accordance with the extant literature, technologies that are cognitively close to a region’s existing patent portfolio enjoy higher patenting activity, confirming that relatedness remains a strong and persistent predictor of inventive output. Second, local AI endowments are positively associated with patenting across technological fields, even after conditioning on relatedness, indicating that AI plays an enabling and cross-cutting role in a given regional innovation system. Moreover, the interaction between relatedness and AI turns out to be negative and statistically significant, implying that AI attenuates the extent to which local inventive activity depends on the technology’s proximity to the regional portfolio. In sum, AI appears to raise the overall local inventive activity and partially relaxes its dependence on relatedness.| File | Dimensione | Formato | |
|---|---|---|---|
|
tesiphd_completa_dalessandro.pdf
accesso solo da BNCF e BNCR
Licenza:
Tutti i diritti riservati
Dimensione
8.99 MB
Formato
Adobe PDF
|
8.99 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/362386
URN:NBN:IT:UNICATT-362386